ai agent governance audit

The problem

Most AI governance is just a policy binder that ignores how agents actually behave. When an agent holds API keys, makes purchases, or deletes data, a "system prompt" is not a security control, and an unproven kill switch is a liability.

What it does

• Audits deployed agents against seven critical runtime controls, including containment, permission scope, and blast radius.
• Differentiates between PASS, FAIL, and UNVERIFIED status based on evidence rather than architectural claims.
• Identifies "decorative" human oversight where approval fatigue renders the human-in-the-loop ineffective.
• Produces audit-ready findings mapped to ISO 42001, NIST AI RMF, and OWASP Agentic AI Top Ten.
• Generates interview scripts and demonstration lists to force teams to prove their safety claims.

Why this beats prompting it yourself

General LLMs treat system prompts as enforcement and often hallucinate framework control IDs. This skill enforces a "test, don't trust" logic that prioritizes exploitability and runtime reality over paperwork, ensuring you don't submit fabricated citations to a real auditor.

Use cases

• Pre-deployment security reviews for agents with write-access to production databases or financial tools.
• Post-incident reconstruction to determine how and why an agent performed an unintended action.
• Preparing evidence and artifacts for ISO 42001 certification or NIST framework compliance.
• Validating that multi-agent systems have sufficient circuit breakers to prevent cascading failures.

Known limitations

This is not a certification or attestation. It does not cover ML model bias or fairness auditing. It requires manual input of architectural details or interview responses to function.