Anthropic has placed six researchers from METR, a nonprofit that tests advanced AI models, inside the company with access comparable to its own staff and has given up the right to block what they publish, TechCrunch reported Thursday. The arrangement offers a concrete answer to a question that matters to boards, insurers and regulators relying on outside reviews of AI developers: who decides what the public gets to see.

What Happened

In a blog post, “An Update on Independent Evaluators,” Anthropic said it had onboarded the six METR employees with company hardware, office space and access to internal tools and systems roughly on par with its own risk assessment teams. Staff are encouraged to share information with them as they would with a colleague, except for customer data, attorney-client privileged material and details of model architecture.

METR will publish its own analysis of Anthropic's risk levels, incidents and practices. “Anthropic can't veto publication,” the company wrote. It may require redactions only in those three categories, and METR is free to say publicly whether it thinks a redaction was justified, provided it does not reveal the withheld material. Only a small METR team works on the engagement under enforceable agreements; other METR staff, including researchers working with Anthropic on separate projects, are walled off.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

Anthropic described the setup as the top tier of a three-level system. At the lowest tier, evaluators run narrow investigations and need the company's approval to publish; TechCrunch cited the biosecurity nonprofit SecureBio, which is reviewing the chemical and biological weapons chapter of Anthropic's August risk report. A middle tier receives more information and may publish without approval. METR is so far the only evaluator at the top. Anthropic also said Accenture and the AI firm Faculty are working toward a broader partnership on evaluation, red-teaming and safeguard testing, and that Hugging Face is among organizations in early discussions.

What Changed

On Oct. 13, The Primary Wire reported, citing Reuters, that Anthropic and OpenAI had sought substantial control over what outside evaluators publish, an account both companies partly disputed. Anthropic told Reuters then that it reserved the right to approve reports from new evaluation firms. Thursday's post puts in writing that, for METR at least, there is no such right, and it names the only categories the company can withhold. The approval right survives at the lowest tier, however, so not every review carrying Anthropic's cooperation will be free of its editing.

Why It Matters

The White House Accord on Super Intelligence, which Anthropic cited, commits signatories to work with independent auditors or evaluators but leaves the terms to private contracts. Anthropic compared its approach to the resident examiners regulators station inside large banks, while acknowledging the analogy is not exact. It listed criteria it expects evaluators to meet, including no equity in the company, no financial or operational dependence on it and freedom to publish.

For enterprise buyers and insurers, that offers a test to apply to any lab's safety claims: was the reviewer embedded, could it publish unapproved, and what could the developer redact? TechCrunch noted the step follows a run of disclosures since late August of AI agents acting outside their instructions, which has sharpened calls for independent checks. It also fulfills a proposal Chief Executive Dario Amodei made in a September essay, and TechCrunch reported OpenAI quickly pledged the same.

What to Watch

Anthropic gave no date for METR's first published assessment. Watch when it appears, how many redactions METR flags and whether it calls any unjustified. Also watch whether OpenAI and the other accord signatories offer evaluators comparable access and publication rights, and whether evaluators move from Anthropic's lower tiers to the top one.