OpenAI's Monitoring System Faces Scrutiny from Third-Party Evaluators

The nonprofit organizations METR and Redwood Research have conducted a pilot review of OpenAI’s internal monitoring system, designed to test the company’s commitment to third-party safety evaluations. The exercise was part of OpenAI’s effort to deepen its work with external researchers and ensure that its AI systems align with human values.

As part of this collaboration, METR and Redwood Research were given access to OpenAI’s internal codebase, documentation, and technical staff from July 21 to August 11. During this period, the evaluators discovered several novel vulnerabilities in the monitoring system, including an exploit that could prevent hundreds of recent actions from being monitored.

OpenAI has since patched these identified vulnerabilities, which impact its production safeguards. The company’s spokesperson stated that they are using what was learned during the pilot to strengthen their safeguards and shape future collaborations with independent evaluators. This commitment is part of OpenAI’s ongoing effort to pioneer new approaches to third-party testing.

The pilot review was conducted in response to a proposal by Anthropic CEO Dario Amodei, who suggested granting employee-like access to independent evaluators as part of slowing the pace of AI development. As a result, OpenAI has pledged to grant such access and is actively finalizing contracts with third-party safety assessors.

President Donald Trump’s voluntary accord for AI companies to partner with external auditors or evaluators has also gained traction in the industry. Anthropic, Google, Meta, OpenAI, SpaceX, and Nvidia have all signed this agreement, which aims to promote transparency and accountability within the multitrillion-dollar AI sector.