OpenAI is letting outsiders look under the hood of its most powerful AI systems — and asking them to check whether the models can secretly work against the people who built them.

The company announced on Monday that it has signed agreements with four independent research groups — Apollo Research, METR, Transluce and Goodfire — for the first phase of what it calls "embedded assessments." The evaluators will get deep access to internal models, training data, source code and monitoring tools, and will work directly alongside OpenAI's own engineers.

It is the most open a frontier AI lab has ever been with outside inspectors, and it comes as the industry faces growing questions about whether increasingly capable models can be trusted to behave as intended.

Apollo's assignment: hunting for subversion

The most eye-catching of the four pilots belongs to Apollo Research, which will place six researchers inside OpenAI for four months to test whether its models can covertly act against their developers' intentions — a risk researchers call subversion.

Apollo will test four claims: that training did not push models toward subversive behaviour, that OpenAI's evaluations and red-teaming would catch it if it happened, that models have not attempted it during internal use, and that any reasoning behind such behaviour would be detectable.

The group plans to publish two public reports on its findings, while unredacted versions go to the Safety and Security Committee of the OpenAI Foundation's board.

What the other three labs will do

METR will run two assessments: an alignment review of what OpenAI describes as its largest reinforcement-learning run to date, and a stress test of the company's monitoring systems following a three-week red-team trial.

Transluce will examine the quality and coverage of OpenAI's post-training alignment evaluations, with an eight-week follow-on assessment planned for early 2027 that expands its access to more models and information.

Goodfire will embed interpretability researchers inside the company to monitor training runs, debug models and assess alignment as the work progresses.

Paid, unpaid — and the right to redact

The money arrangements split the four groups. Apollo Research and Goodfire accept payment from OpenAI, though the company says it is not tied to their findings. METR and Transluce take no fee, but may accept computing power or API credits to support their work.

The evaluators will be able to publish what they find — but OpenAI reserves the right to redact confidential information. Apollo says it will explain when redactions limit what readers can verify, and OpenAI says it cannot withhold information simply because it is unfavourable.

The programme follows a pledge by chief executive Sam Altman in September to give independent evaluators deep access to the company's work, and supports OpenAI's commitments under the White House Accord on Super Intelligence, which calls for independent verification of safety controls.

OpenAI acknowledges that norms for this kind of embedded assessment have yet to be standardised, and expects technical access and security issues to arise. For now, the pilots are a first step — the results are not yet available, and the real test will be whether findings that embarrass the company ever see daylight.