AI agent audit trails: detect risky tool use, leaked secrets, prompt injection, and missing human approval.