Encased — AI agent security
← Blog

Guide · 2026-10-19

AI Agent Security Checklist: 12 Controls Before Production

A practical AI agent security checklist covering credentials, network egress, prompt injection, spend limits, human approval, logging and incident response.

How to use this checklist

Work through it per agent, not per platform. An agent that only summarises documents needs fewer controls than one that can issue refunds. Mark each item done, not applicable (with a reason), or open.

Identity and access

1. Each agent has its own credentials, scoped to the minimum tools it needs.

2. Credentials can be revoked instantly without redeploying.

3. Model-provider keys are never exposed to untrusted agent code.

Inputs and actions

4. Untrusted content (email, web, tickets, PRs) is treated as data, never as instructions.

5. Outbound network destinations are restricted to an explicit allow-list.

6. Destructive or irreversible actions require human approval of the exact action.

Runtime limits

7. Per-agent token and spend caps are enforced before calls, not just reported after.

8. Recursion depth, step counts and wall-clock time are capped.

9. Every control fails closed when the policy service is unreachable.

Evidence and response

10. Every proposed action and verdict is logged with timestamps, redacting secrets and personal data.

11. There is a tested emergency stop and a named person who can use it.

12. Policies are versioned and reviewed when agents gain new tools.

Next step

If several items are open, a focused AI agent security audit can map the gaps and produce a policy blueprint. Encased's audit covers up to ten agents over five days; it is not a penetration test or certification.