Proposed agent audit checker reportedly proved 42.2% of tested rule-breaking runs

A post describing a Carnegie Mellon paper says its checker proved 42.2% of 696 rule-breaking runs across three benchmarks. The proposed “verifiable agent insurance” uses a provider’s blockchain deposit to pay users when a verifier confirms rule breaches using tool-call logs and on-chain hashes. Approvals buried in chat text often blocked proof; the authors estimate structured approval records could raise the figure to 83.0%, with under 1¢ in blockchain fees per session.