METR and Redwood red-team OpenAI's internal agent monitoring and find ways agents could avoid or disable it

METR · Research report

Read the source (opens in a new tab)
What happened
METR said on October 22, 2026 that a METR and a Redwood Research employee spent three weeks in July and August testing OpenAI's internal misalignment monitoring, with deep access.
Key points
  • They estimate roughly 75 to 94% of internal agentic Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring; RL training runs were not covered.
  • Not monitoring compaction events likely left on the order of 10% of eligible internal Codex traffic unmonitored from May to July 2026; OpenAI had found and fixed this.
  • They found a way Codex agents could bypass sandboxing and Auto-Review, an exploit hiding hundreds of recent actions from the monitor, and two headers that disable monitoring for new agents.
  • They say future monitoring should withstand "deliberate, strategic, and sustained attacks from highly capable internal agents at scale."
Evidence
External red-team findings on the system as it stood on and before July 21. METR publishes a small subset of a 61-page internal report, and OpenAI could redact this post.
Uncertain
OpenAI says all vulnerabilities affecting its production safeguards are patched and coverage has changed; the authors did not evaluate these claims.

See it in the latest record