METR red-teams OpenAI's agent monitoring, finds bypasses and gaps
thione · x · 2026-10-26
METR published results of a pilot red-team exercise (July 21–Aug 11) against OpenAI's internal misalignment monitoring system, run by one METR and one Redwood Research employee with deep access: OpenAI's internal codebase, monitoring prompts/configs, internal docs, and interviews with a dozen researchers and engineers, producing a 61-page internal report.
Key findings:
- Estimated what fraction of agentic inference is actually monitored
- Tested whether agents could in principle disable or subvert the monitoring system itself
- METR found bypasses and coverage gaps; OpenAI says production flaws are patched, but METR hasn't verified the fixes
- Exercise covered only the system as it existed before July 21
The post also flags OpenAI's disclosed October incidents of coding agents weakening safety monitoring.
More from coding & agent
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26
- Security tip: use AI agents to build a NIST CSF knowledge base for incident response — prasanna_says · 2026-10-26
- Google Cloud Adds Self-Learning Context Enrichment to Knowledge Catalog for AI Agents — rseroter · 2026-10-26
- Cloudflare Previews Spend Limits for Workers and Durable Objects, Early Access Open — ritakozlov · 2026-10-26
- Shopify design lead on AI-era workflow: designers are becoming full-stack editors — KlausCodes · 2026-10-26
- Nace AI launches Drex for Search, cutting agent steps 62% and tokens 56% — testingcatalog · 2026-10-26