External Evaluators Uncover Flaws in OpenAI's Internal Monitoring During Trial Run

Two nonprofit research groups say they found weaknesses in OpenAI's internal monitoring setup during a trial program that gave outsiders a close look at the company's systems.

METR and Redwood Research said the exercise, which ran from July 21 to Aug. 11, was meant to show what it would look like to hand external researchers deep access to OpenAI's infrastructure. One staffer from each organization could see OpenAI's internal codebase, documentation and speak with some technical employees. They were supervised on site and used company-issued devices that never left the building, according to a release.

The reviewers turned up several new vulnerabilities, including one that would let agents keep hundreds of recent actions from being monitored. OpenAI told the groups it has fixed the vulnerabilities affecting its production safeguards.

The two organizations said they would publish only a high-level summary, since the pilot wasn't built to give the public a full picture of OpenAI's monitoring. They gave OpenAI a 61-page internal report and expect to share more about future exercises.

An OpenAI spokesperson said the work builds on its efforts to pioneer new approaches to third-party testing, and that the company is using what it learned to strengthen safeguards and shape future collaborations with independent evaluators.

The pilot follows CEO Sam Altman's September pledge to give independent evaluators employee-like access, an idea Anthropic CEO Dario Amodei proposed as part of his plan to slow AI development. OpenAI said earlier this month it is finalizing contracts with third-party safety assessors.

President Donald Trump, who has favored limited AI regulation, has backed outside assessment, urging AI companies to partner with an independent external auditor or evaluator under a voluntary accord presented at the White House. Executives from Anthropic, Google, Meta, OpenAI, SpaceX and Nvidia signed it.

Experts told CNBC that questions remain about how such evaluator groups will be funded and how their reporting will be structured.

Source: CNBC