Carnegie Mellon paper finds agent logs prove only 42.2% of policy violations

A Carnegie Mellon paper, Holding Agents Accountable with Verifiable Agent Insurance, finds that agent logs often cannot prove policy violations.

Across 696 rule-breaking runs from three benchmarks, the authors' checker proved 42.2% of violations, often failing where approvals appeared only in chat messages.

The authors estimate that structured approval records would raise the provable share to 83.0%, at under 1 cent of blockchain fees per session. This estimate is the authors' own projection, not a measured result.

A social media post by Rohan Paul reported the findings, and one summary of the post described OpenClaw's logs as leaving 0 of 4 policy violations provable.