AI Agents Policies
AI Agent Policies
An AI agent that is unowned, over-privileged, or running on a shared or static credential is one of the fastest routes to your data, because it acts on its own, at machine speed, with valid credentials that nobody is watching. AI Agent Policies encode what good agent posture looks like, so every agent in your environment is judged against the same bar instead of relying on manual review.
Available AI Agent Policies
Identity & Secrets Intelligence currently ships the following built-in AI Agent Policies:
| Policy | Severity | What It Flags |
|---|---|---|
| Agentic Privilege Escalation | Critical | An agent whose execution role holds more privilege than the identity that created it |
| Orphaned Owner | Critical | An active agent whose attributed owner or creator no longer resolves in IAM |
| Insecure Inter-Agent Trust | High | Two or more collaborator agents sharing a single execution role instead of independent identities |
| Supervisor Aggregation Excess | High | A supervisor or router agent aggregating access through more than 5 collaborator agents |
| Missing Guardrail | High | An agent with no guardrail attached to filter harmful content or denied topics |
| Session/Credential Inheritance | High | An execution role that can mint a human-scoped workload token instead of the agent's own identity |
| Static/Long-Lived Credential in Use | High | A static, key-shaped credential in an action-group function's environment variables |
| Shared Credential Across Agents | High | The same static credential referenced by two or more distinct agents |
| Credential Age Violation | High | Automatic rotation disabled on the agent's customer-managed KMS key |
| Unowned Agent | Medium | An agent with no owner tag and no creator attributable from CloudTrail |
| Missing Agent Decision Record | Medium | An active agent with model-invocation logging disabled, leaving no record of its decisions |
| Retention Non-Compliance | Medium | Agent invocation logs retained for fewer than 90 days |
| Wildcard Resource Binding | Medium | An action-group function whose resource policy is not pinned to the invoking agent |
| Cross-Account/Cross-Tenant Reach | Medium | An agent's execution role or tool resources reachable from outside the account or organization |
| Decommission Overdue | Medium | An agent flagged stale that is still active more than 180 days after its last change |
What's Next
Updated 7 days ago
Did this page help you?