CMU paper: agent logs prove 0 of 4 violations; structured records raise detection to 83%
rohanpaul_ai · x · 2026-10-14
A Carnegie Mellon paper, "Holding Agents Accountable with Verifiable Agent Insurance," tackles what happens after an agent misbehaves rather than just prevention.
Key findings
- Analyzing OpenClaw logs, 0 of 4 policy violations could be proven from logs, since approvals and permissions are buried in chat text rather than emitted as structured data.
- Only 10-30% of violating runs actually need human judgment—most agent rule breaks could be checked by code with better logs.
Verifiable Agent Insurance (VAI)
- Providers post a deposit on an Ethereum L2 with programmatically checkable promises; a neutral verifier compares tool-call logs against on-chain hashes, and the deposit pays users when rules are broken.
- On 696 violating runs across 3 benchmarks, VAI proved 42.2% of violations; with structured approval records, detection rises to 83.0% at under 1¢ in blockchain fees per session.
Practical takeaway: design agents so every permission, its scope, and each risky action lands in machine-readable logs.
Related event: CMU Paper: Structured Logs Could Make AI Agent Violations Provable(2 posts)→
More from coding & agent
- Harness Launches a $149 Command Center to Run All Your AI Agents — dee_hw · 2026-10-14
- Dev runs Codex as a shell tool inside Claude for code review, skipping tmux hassles — abhi9u · 2026-10-14
- Using AI agents turns out to be accidental management training — dr_alphalyrae · 2026-10-14
- Rerunning the single-task experiment across all ipython harness versions produced the benchmark plot — yb2698 · 2026-10-14
- Small-model agent harnesses are fragile to environment noise — open-source ipython harness released — yb2698 · 2026-10-14
- For small models, ipython is all you need: single-tool harness ties production-grade scaffolds on TB-Lite — yb2698 · 2026-10-14