Security

Read-only by design

You create the role yourself, from a template you can read before you create it. The role's trust requires an ExternalId that is unique to your connection.

What it reads, and what it can't

Watch scans read the same read-only data as the nightly scan, every 15 minutes, and never change anything.

The role can read only what OpsAgent's checks use:

  • Compute and network: your instances, volumes, snapshots, Elastic IP addresses, security groups and network interfaces, your load balancers, and the list of regions.
  • Databases: database instances and clusters, their snapshots, and who each RDS instance snapshot is shared with.
  • Storage: bucket names, their regions, their public-access settings and whether a bucket's policy makes it public.
  • Identity: the credential report (which users have a password, MFA and access keys, and when each was last used; never the passwords or keys themselves), the policies attached to users, and whether the roles your backups use are allowed to restore.
  • Logging and threat detection: your CloudTrail trails and whether they are logging, whether GuardDuty is on, and log group names and settings.
  • Backups and certificates: backup plans, what they cover, backup jobs and vaults; certificates and their expiry dates.
  • Usage and cost: CloudWatch metrics such as CPU and GPU use, the list of SageMaker training jobs, and your daily cost per service.
  • Itself: the account ID, and its own role and policies, so OpsAgent can check the role still matches the template.

It cannot read object contents, database rows, secret or parameter values, log events or function code. The role does not grant any of these, and it explicitly denies reading objects, secrets, parameters, table items and function code. The role has no write permissions at all, so OpsAgent cannot change, stop or delete anything.

One read is optional: Terraform state files. The template has a setting that lets the role read state files under one bucket and prefix you name. It is off unless you fill it in, and OpsAgent does not read state files today. State files can contain secrets, so name only the prefix you want checked.

Each full scan (the nightly one, or one you start yourself) makes one AWS Cost Explorer request in your account, which AWS bills at about US$0.01. Watch scans make none.

Your data

OpsAgent stores inventory metadata, findings and the evidence behind them. It is stored on Amazon Web Services, in the United States (Oregon).

OpsAgent's outbound requests, such as calls to your own AI endpoint, come from one fixed address, 32.185.67.24, in case you allowlist by IP. Requests to Amazon S3 in the Oregon region travel inside AWS and do not use it.

Secrets such as Slack tokens are encrypted with a per-organisation key, which is itself wrapped by a master key.

AI and your data

Dashboard AI is off until an owner turns it on. Only then does OpsAgent send the AI the findings and resources it looks up to answer you, which can include lists across your organisation. They go to Anthropic, to OpenAI as the backup provider, or to the AI provider or endpoint an owner configures. Anthropic does not train on this data.

Agents you connect over MCP use their own model. MCP traffic never touches OpsAgent's model key.

If you add your own key, it is encrypted with your organisation's key, and only its last 4 characters are ever shown again.

You stay in control

Delete the role at any time. Disconnecting stops OpsAgent from using the role at once. Deleting the stack in your account removes the role entirely.

Reporting a security issue

Use the contact form and choose Security.