METR and Redwood Found Six Ways to Blind OpenAI's Coding Agent Monitor

During a three-week embedded pilot at OpenAI, researchers from METR and Redwood Research found multiple exploits that let AI agents evade the internal monitors meant to catch misalignment. A brief write-up of their work is now public. Why it matters: the result shows that even a frontier lab's own oversight tooling can be blinded — raising the bar for independent evaluation of agent monitoring before these systems are deployed more widely.

• Alpha Signal2026-10-22 15:26
• OpenAI2026-10-22 15:57