Agentix Labs security resource

Agentic AI SOC 2 Security Review Checklist

Agentic AI introduces new control questions for SOC 2 teams: what can the agent access, what can it change, who approves risky actions, and how do you prove what happened?

Access controls

  • List every tool, connector, token, and MCP server.
  • Use least-privilege access by workflow.
  • Rotate and revoke secrets through managed systems.
  • Separate human and agent credentials where possible.

Change control

  • Require approval for high-risk writes.
  • Log prompt, context, tool call, output, and approver.
  • Keep rollback steps for partial or failed actions.
  • Review production changes in a normal change-management flow.

Monitoring

  • Alert on failed writes, unusual volume, repeated retries, and permission denials.
  • Retain run logs long enough for audit review.
  • Map incidents to owners and response actions.
  • Measure false positives and false negatives.

Vendor and data risk

  • Document data sent to model providers.
  • Review retention, training, residency, and subcontractor policies.
  • Mask or avoid sensitive fields when not needed.
  • Keep a vendor evidence record.

How to score it

Give one point for every checked item. Then use the result to decide what happens next.

Need a security review for an agentic workflow?

Book an agent security teardown