Summary
The authors show how every individual agent step can pass its own check while the complete execution violates authority, threshold or review requirements. They propose provenance-aware monitoring that recomputes guarded quantities across the full execution trace rather than relying only on per-step classifiers.
Why it matters
The result highlights a structural weakness in current agent governance: local compliance does not guarantee that an autonomous workflow remains within its intended authority.
Related coverage
- OpenAI introduces a model-misalignment disclosure framework and six case reports
- Google Home opens device control to third-party AI agents
- A new paper models collective loss of control in agent swarms