OpenAI said on Monday, Nov. 16, that it has signed agreements with four outside groups that study AI risks, Apollo Research, METR, Transluce and Goodfire, to test the safety of its AI from inside the company. OpenAI says that, depending on what each review needs, the groups get some of the same internal resources its own researchers use. That includes versions of models saved partway through training, source code, training and test data, monitoring systems, and talks with the staff who did the work. The deals build on earlier work. METR recently spent three weeks testing the systems OpenAI uses to watch its AI, with access to its code and staff, and Apollo has tested OpenAI's models alongside the company before.
The first round looks at whether OpenAI's models do what the company intends, and whether its safeguards hold, while the models are being trained and used inside OpenAI. Apollo says its own earlier testing of leading models came after the final stage of training and shortly before public release. It says that is not enough to test for hidden behavior that goes against what a model's makers intended, or to check how models are used privately inside a company.
Each group has its own job. Apollo will send six people to work inside OpenAI for four months, checking whether models could covertly work against what their developers intend and whether OpenAI's safeguards would catch it. METR will look for behavior OpenAI didn't intend in its largest reinforcement learning run to date. Reinforcement learning is learning by trial and error, with a reward for getting it right. OpenAI says this scope is narrower than the broad review METR proposed. The search will include failures OpenAI's own tests and monitors may have missed. Transluce will review how good and how thorough OpenAI's tests are for checking whether its models behave as intended. OpenAI and Transluce say they intend to follow up with a review of about eight weeks in early 2027. Goodfire will study what goes on inside the models, mainly to help OpenAI's own researchers, and OpenAI says Goodfire will also publish what it learns as the work goes on.
The groups can publish their findings, but OpenAI keeps the right to remove confidential information, and it may ask them to wait before publishing so it can fix a safety or security problem first. Apollo says OpenAI agreed not to hold back information, or ask for anything to be cut, just because a finding looks bad, and OpenAI cannot cut Apollo's verdict on each question it tests. OpenAI can still remove supporting evidence it considers sensitive intellectual property, so readers may not see everything behind a conclusion, though Apollo says it can say whether those cuts affected its conclusions.
Talks with OpenAI staff are off the record: a group can publish what it learned in them only if OpenAI formally confirms it or the group shows it on its own, and it can say when OpenAI declined to confirm something. Apollo also says it may discuss confidential matters only with staff members OpenAI picks, though it can suggest people. Under Apollo's terms, OpenAI may hold back some information within the agreed scope for security, intellectual property or technical reasons, but it must say so and give written reasons. Apollo adds that the pilot does not yet give it all the access and freedom to publish it believes a full outside review needs, and that only the access for the first question it tests is settled, with the arrangements for the rest still to be agreed.
OpenAI also set out who pays, who settles disputes and where the work happens. Apollo and Goodfire are paid for their work, though OpenAI says their pay does not depend on what they find. OpenAI says METR and Transluce take no fees but may accept computing time. Disagreements between OpenAI and a group go first to an internal OpenAI safety panel and company executives. If they remain, they go to the Safety and Security Committee of the OpenAI Foundation's board, which OpenAI calls independent. That committee will see the full, uncut reports. Where access could expose its most sensitive work, such as the models' inner design, OpenAI says it may have its own researchers run analyses with the group instead. Group members who handle sensitive information work in OpenAI's offices on OpenAI devices, which OpenAI says is to protect that information.
OpenAI says the deals help meet its promise, under the White House Accord on Super Intelligence, a pledge AI companies signed to work with outside reviewers, to have its safety controls assessed independently. Rival AI company Anthropic said in October that it had begun giving a six-person METR team what it calls "employee-like access." Apollo plans two public reports, but neither it nor OpenAI has said when they will come out. A second METR review of how OpenAI watches its AI, which OpenAI intends to be broader than that three-week test, is still waiting on the two sides to agree on its scope.