GoKawiilMETR, Redwood Research find flaws in OpenAI's internal monitoring system during pilot audit
Nonprofit AI safety evaluators METR and Redwood Research spent three weeks in July and August with employee-level access to OpenAI's codebase, documentation and staff to test an internal monitoring system. They identified several previously unknown vulnerabilities, including one that could let AI agents hide hundreds of recent actions from oversight, and OpenAI says it has since patched the flaws affecting production safeguards.
GoKawiil's interpretation of the reporting above, not reported fact.
The exercise reflects a broader push, prompted partly by Anthropic CEO Dario Amodei's call for slowing AI development, to give outside researchers deeper access to frontier AI systems for safety checks. The evaluators' decision to withhold a detailed public report suggests current third-party testing arrangements still fall short of full transparency, even as companies tout them as evidence of safety commitments.
- METR and Redwood Research had employee-like access to OpenAI's systems from July 21 to Aug. 11.
- They found novel vulnerabilities, including one letting agents evade monitoring for hundreds of actions, which OpenAI says it fixed.
- Only a high-level summary was made public; a 61-page detailed report was shared privately with OpenAI.
Source: cnbc.com — Ashley Capoot, 2026-10-22
Published there as: “Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.