Anthropic has placed six employees of METR, a nonprofit that tests the most powerful AI models, inside its own operation, with the right to publish their conclusions about the company’s risks. Anthropic announced the arrangement on 15 October. Most safety claims in the industry are still graded by the companies making them, so an outside team with real access and the freedom to say what it sees is a different kind of check.
The details decide whether it is. Start with what “employee-like access” covers. According to Anthropic’s post, the METR staff have company hardware, office space, and internal tools and systems roughly on par with what Anthropic’s own risk assessment teams use. Anthropic staff are told to treat the visitors like a colleague on the alignment-risk side and brief them freely. The post also says evaluators need visibility into how models were trained and how safeguards work.
The exclusions are named too. Customer data, attorney-client privileged material, and information about model architectures sit outside the default, and the post calls these “a small number of clearly defined categories.” The post does not mention model weights or pre-release checkpoints, so the reader cannot tell from it whether those fall inside or outside the line. Anthropic also says it cannot publicly share some of its own material, including compute and training information, because it is commercially sensitive. Whether METR sees that material is a question the announcement leaves open.
Publication is the part Anthropic answers most directly. It cannot veto what METR publishes. Anthropic can demand redactions, but only where material falls inside a short list of predefined categories of the same kind as those exclusions. METR may tell the public that cuts were made and say whether it thinks they were justified, so long as it does not reveal what was removed. The post does not say whether Anthropic sees drafts before release, whether there is a notice period, or who rules on whether a redaction fits a category. Those mechanics determine how much of the right to publish survives contact with a product launch.
Funding is the gap that matters most. Anthropic lists independence as a requirement for any evaluator: no equity in the company, and no financial or operational dependence on it. The post does not say whether METR is paid by Anthropic, or by whom, so the reader is told the standard but not whether this arrangement meets it. The post does say METR has worked with Anthropic on reviews of its Risk Reports over the past year, and that other METR researchers work with the company on separate projects. Anthropic describes those researchers as walled off from the embedded team.
The word “independent” can blur a distinction here. METR’s judgment may be its own while its access, and possibly its funding, still depend on Anthropic. Access is granted by Anthropic and can be withdrawn by it. The post says the engagement runs under enforceable agreements and is voluntary. It does not say whether the access carries over to future models automatically or is renewed model by model.
METR is the top of a three-tier scheme. At the first tier, Anthropic decides what an evaluator sees and publication needs Anthropic’s approval. At the second, evaluators publish without approval and see more, though Anthropic can still demand removal of specific categories. The third, METR’s, adds employee-like access with second-tier publication rights. Anthropic says it is also working with Accenture and Faculty, and is in early talks with others, Hugging Face among them. SecureBio is partway through reviewing the August Risk Report section on chemical and biological weapons. Those partners get less information and fewer publication rights for now. The stated aim is to earn trust and move them up.
Anthropic compares the model to bank supervision, where regulators station examiners inside the largest banks. The analogy is imperfect, as the post admits, and one difference is plain: examiners answer to a regulator with legal force, while METR’s seat exists because Anthropic chose to offer it. The company says it believes every frontier AI company should embed evaluators the same way. The post names no other lab that has.
The first test will be METR’s first published assessment. If it includes a stated redaction that METR disputes, the arrangement is working as described. If it arrives without one, readers will have no way to tell whether nothing was withheld or nobody objected.
Based on Anthropic’s newsroom post “An update on independent evaluators” of 15 October 2026
