METR, Redwood Research find exploit in OpenAI agent monitoring that could hide hundreds of actions

Overview

METR and Redwood Research reported on October 22, 2026 that a three-week embedded pilot at OpenAI identified previously unknown weaknesses in the company’s internal monitoring of AI agents. The groups said one exploit could have enabled agents to stop hundreds of recent actions from being recorded by the monitoring system. OpenAI said it has patched the identified vulnerabilities affecting its production safeguards. The pilot ran from July 21 through August 11, 2026. One employee from each nonprofit received high-level access to OpenAI’s internal codebase, monitoring prompts and configurations, documentation, and interviews with roughly a dozen staff members from monitoring, safety, and security teams. Their work included estimating monitoring coverage for agentic inference traffic through the Responses API and OpenAI’s internal use of Codex. The evaluators estimated that monitors covered about 75% to 94% of Responses API traffic during the period, leaving some traffic unmonitored. They delivered a 61-page internal report to OpenAI researchers and security staff; OpenAI-redacted a 38-page version for the leadership of METR and Redwood Research. The organizations said the pilot was not designed to provide robust public transparency, so they agreed to publish only high-level findings. The exercise followed CEO Sam Altman’s pledge to give independent evaluators employee-like access and occurred while OpenAI was finalizing contracts with third-party safety assessors.

Summary from vectorwire.ai.