AgentGavel Puts AI Agent Governance Under Attack

Field notes

AgentGavel structures each test around a setup, an adversarial probe, observable events, and a deterministic validation rule. Its methodology separates runtime-enforced controls from model refusals, treating the latter as a repeated-run rate rather than proof of enforcement. The documented cases include replayed approvals, undeclared tool parameters, secrets entering model context, and audit records whose order or contents have been altered.