For CISOsLast reviewed

Checklist for CISOs and audit evidence

Ask these questions of any vendor, or of the internal team proposing to build it themselves. They follow the same six criteria as the ranking, so the answers map straight onto the scores. Then check the program produces the evidence below.

Questions to ask

01

Visibility

Can you see which agents, MCP servers and skills run, on which machines, and what they do?

  • Show me every coding agent, IDE extension and MCP server running on our machines today, and who installed each.
  • How do you find an agent or MCP server we have not approved?
  • Can you see the autonomy settings each developer has chosen, such as auto-approve?
02

Enforcement

Can it stop or hold a risky action before it runs, not just report it after?

  • Where exactly is policy evaluated: before the action runs, or after?
  • Which outcomes can a policy have: allow, warn, block, require approval?
  • How does a rule tell a destructive command from a routine one, beyond matching its text?
  • Can one policy apply to every agent we use, including the next one we adopt?
03

Developer friction

Do developers keep their tools and flow, or do they look for a way around the control?

  • What does a developer see when an action is blocked or held, and how long does an approval take?
  • Do developers keep their current agents and settings?
  • How do you keep prompts rare enough that people still read them?
04

Coverage of terminal and MCP actions

Does it understand shell commands and MCP tool calls as agent actions?

  • Do you see terminal commands and MCP tool calls, for local and remote servers?
  • Which agents are covered today, and on which operating systems?
  • Do you see what agents read from files and send to model providers?
05

Audit evidence

Can you show an auditor what agents did, under which policy, and who approved it?

  • What is recorded for each action: agent, user, command or tool, policy, decision, approver?
  • How long are records kept, and can they go to our SIEM?
  • Can you produce the record for one developer's sessions on one day?
06

Deployment effort

How much work is it to roll out across every developer machine and keep running?

  • How is it deployed to every machine, and what does our MDM need to do?
  • What does day two look like: who writes and tunes policy, and how long does that take?
  • What happens if the control is unavailable: do agents fail open or closed?

The audit evidence an AI agent program should produce

If an auditor or a customer asks how AI agents are governed, these are the records that answer.

  1. Agent and MCP inventoryEvery agent, extension and MCP server per device and user, with first-seen dates and approval status.
  2. Written agent policyWhat agents may do, by team and environment, with the owner and the date it was last reviewed.
  3. Decision logEach blocked, warned and approved action with agent, user, policy and time, retained for your audit period.
  4. Exception approvalsWho approved each held action or exception, when, and why.
  5. Configuration baselineApproved agent settings and autonomy levels, and evidence that devices match them.
  6. MCP server reviewsThe checklist result for each allowed server and version, and the date of its next review.
  7. Credential scope for agentsWhich credentials agents can reach, and the scoping or short-lived tokens that limit them.
  8. Incident recordsAgent-related incidents and near misses, what the logs showed, and what changed after.
  9. Periodic access reviewA recorded review of agents, servers and policies, with removals of what is no longer used.

Getting the first three in place

The inventory, the policy and the decision log are what an agent access security broker produces by design. Unbound offers a free two-week assessment for teams with 50 or more developers using coding agents: every agent, MCP server and skill on the fleet, and which actions to stop.