METR and Redwood’s July–August pilot found gaps and bypasses in OpenAI’s agent monitoring; OpenAI says production-safeguard flaws are patched, but the assessors have not evaluated those claims.

METR’s October 22, 2026 report says OpenAI claims the identified production-safeguard vulnerabilities are patched and explicitly says the assessors have not evaluated those claims.

Published 22 Oct 2026 · 16:40 UTC

METR and Redwood Research ↗ · Source dated 22 Oct 2026

Browse all Live updates · Read Features