Is the list of actions each AI agent can perform limited to what its task needs, and written down?
Facts read 16 Nov 2026. Page reviewed 16 Nov 2026.
Question 24 of the assessment asks: Is the list of actions each AI agent can perform limited to what its task needs, and written down? An agent that can call any API will use it when it is tricked or wrong, so the limit has to sit in the permissions, not in the prompt.
Start the assessment ยท All 40 questions
Why it matters
An agent that can call any API will use it when it is tricked or wrong, so the limit has to sit in the permissions, not in the prompt.
Answers and what each is worth
| Answer | Value |
|---|---|
| Yes, scoped and reviewed | 1 |
| Partly scoped | 0.5 |
| No or not known | 0 |
How to check it in each cloud
- AWS
- Review the agent's role policies and the action groups of each Bedrock agent for wildcard actions or resources, using Access Analyzer policy validation.
- Azure
- Review role assignments of each AI app identity and AI Foundry connections for broad roles such as Contributor or Owner.
- Google Cloud
- Review IAM bindings of each Vertex AI agent's service account for primitive or broad roles, using the IAM recommender.
The fix
Replace broad roles on agents with policies naming only the actions, resources and data each agent needs. Effort: high.
Where it comes from
- Microsoft cloud security benchmark AI-4: Apply least privilege for agent functions
- NIST SP 800-53 Rev. 5 AC-6: source
Related pages
Sources
- Microsoft cloud security benchmark AI-4, read 16 Nov 2026
- NIST SP 800-53 Rev. 5 AC-6, read 16 Nov 2026
Last reviewed 16 Nov 2026.