Nonprofit evaluators METR and Redwood Research said on 22 October that they had reviewed OpenAI’s internal monitoring system. They found several novel vulnerabilities, including a route that would let agents keep “hundreds” of their recent actions out of the monitoring.
Three weeks, two people, OpenAI’s devices
From 21 July to 11 August, one employee from each organisation had access to OpenAI’s internal codebase, its documentation and some of its technical staff. They were chaperoned on site and worked on OpenAI-issued devices that never left the premises.
What they stress-tested was the misalignment monitoring system, as it stood on and before 21 July. The evaluators note that OpenAI kept expanding the system during the exercise. OpenAI has told them the identified vulnerabilities affecting its production safeguards are patched.
A 61-page report nobody will read
Only a high-level post went public. The reason is in the release: the pilot was not designed to provide robust public transparency about OpenAI’s monitoring system. OpenAI received the 61-page internal report, and future exercises are meant to share more publicly.
An OpenAI spokesperson told CNBC the collaboration builds on the company’s commitment to pioneering new approaches to third-party testing, and that it is using what it learned to strengthen its safeguards.
The run-up comes from Anthropic
Sam Altman promised independent evaluators “employee-like access” in September, after Dario Amodei put the idea into the debate as part of his plan for a slower pace. Anthropic has since signed agreements with Accenture, METR and SecureBio; I wrote about METR’s access at Anthropic last week — six evaluators with company hardware, desk space and employee-level access. OpenAI said earlier this month that it is finalising contracts with third-party safety assessors.
Donald Trump backs the idea too. In a voluntary accord he presented at the White House, he encouraged AI companies to partner with an independent external auditor or evaluator. Amodei signed, as did executives from Google, Meta, OpenAI, SpaceX and Nvidia.
Three weeks for something that needs months
The finding itself is good news for OpenAI: the evaluators found real gaps and the gaps are closed. The frame is the problem. A three-week exercise with two people, whose result stays unpublished across 61 pages, is a trial run for a process rather than the process.
Experts raised the same open question with CNBC: how do these organisations get funded, and who do they report to? As long as the answer is “by the company being reviewed” and “to the company being reviewed”, the review stays a favour.