OpenAI Patches Monitoring Gaps Found by Outside Reviewers

METR and Redwood Research have completed a pilot review of OpenAI's internal monitoring systems and found vulnerabilities, including a loophole that lets AI agents evade oversight, according to CNBC. OpenAI said it has patched the issues affecting real service safeguards.

The review ran from July 21 to August 11. One employee from each organization examined OpenAI's codebase and internal documents under supervision, using only devices supplied by OpenAI, and spent three weeks testing the company's system for catching failures in AI model alignment.

Among the findings was a workaround that could allow AI agents to escape monitoring and potentially pull hundreds of recent actions out of scrutiny. OpenAI told the two organizations that the flaws tied to real-world service safeguards had been fixed.

METR and Redwood Research opted to publish only summary results, and delivered a 61-page internal report solely to OpenAI. An OpenAI spokesperson said the findings are being used to strengthen safeguards and to shape future work with independent evaluators.

Background

In September 2026, OpenAI CEO Sam Altman promised independent evaluators staff-level access to the company. Independent evaluation groups have had a limited presence in the multi-trillion-dollar AI industry, and experts say their funding and reporting structures remain unclear.