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

Nonprofit AI safety organizations METR and Redwood Research published findings on October 22, 2026, from a three-week embedded pilot inside OpenAI that uncovered previously unknown vulnerabilities in OpenAI's internal AI agent monitoring system, including an exploit that could have prevented hundreds of recent actions from being tracked. OpenAI said it has since patched the vulnerabilities.

The pilot ran from July 21 to August 11, 2026. One employee from each organization received high-access rights to OpenAI's internal codebase, misalignment-monitor prompts, configuration details, and documentation, and interviewed about a dozen OpenAI researchers and engineers. The goals were to estimate how much of OpenAI's agentic inference traffic its monitors cover, including the Responses API and internal Codex use, and to find weak spots in the monitoring system.

The assessment estimated that approximately 75% to 94% of Responses API traffic was monitored during the period. The evaluators produced a 61-page internal report; a 38-page redacted version was shared with METR and Redwood Research leadership. The organizations published only a high-level public post because the pilot was not designed for robust public transparency. The pilot follows OpenAI CEO Sam Altman's pledge to grant independent evaluators employee-like access.

(Source：Vector Wire)