Security review for AI agent systems
Security for an agent system spans the model, tools, data, credentials, policy engine and external services. Review the end-to-end path an action can take; a safe prompt alone cannot constrain a vulnerable integration.
Threat model the workflow
List assets, actors, trust boundaries and potential impact. Include malicious or compromised tools, prompt injection, confused-deputy behaviour, credential theft, over-broad permissions and accidental disclosure.
- Separate user, agent, service and administrator identities.
- Keep secrets out of prompts, logs and untrusted tool output.
- Validate inputs and enforce policy on the server side.
- Limit tool access by environment, action and resource.
Operational controls
Use monitored approvals for high-impact actions, a tested pause or revocation path and an incident process with named owners. Logs should support investigation without collecting more personal or confidential data than necessary.
Test dependencies and failure modes, including retries, timeouts, partial execution and provider changes. Publish certification, audit or compliance statements only when supported by current scope-specific evidence.
No certification claim
This checklist is general information, not a security audit or certification of Noesia. A vendor review should request current evidence, system boundaries, subprocessor details, incident contacts and remediation history.