Anthropic embeds six METR auditors with broad internal access—but keeps defined areas off limits

Six evaluators will work inside Anthropic

Anthropic has given six employees of the nonprofit Model Evaluation and Threat Research, or METR, company hardware, office space and access to internal systems. The company describes the arrangement as “employee-like”: the evaluators should receive information broadly comparable to what Anthropic’s own risk teams can see. The move was first reported by The Decoder after Anthropic published the terms on October 15, 2026.

The purpose is wider than testing a finished Claude model through a public interface. Anthropic says meaningful evaluation may require visibility into training, safeguards, incidents and internal decision-making. METR is expected to assess risk levels, particular events and company practices, then publish its analysis. No findings from this new arrangement have been released yet, and neither organization has set out a public completion date or a fixed list of questions the first report will answer.

Publication is independent, but access is not unlimited

Under Anthropic’s published terms, the company cannot veto a METR report. It can, however, require removal of information in predefined categories. Customer data, attorney-client privileged material and details of model architecture are normally excluded. Anthropic ultimately decides whether information falls within those categories, while METR may state publicly what level of access it received and whether it considered a redaction justified.

The six-person team also works under enforceable agreements and is separated from other METR staff, including colleagues involved in different Anthropic projects. Anthropic presents this as the highest of three access levels it is developing for external evaluators. SecureBio, Accenture and Faculty currently have narrower arrangements, while discussions with Hugging Face are at an early stage.

That structure gives the auditors substantially more context than a conventional “black-box” evaluation, in which researchers can only send prompts and observe outputs. It is still a voluntary company arrangement, not regulatory inspection with compulsory powers. The distinction matters when interpreting any eventual report.

METR’s independence will be judged through its disclosures

METR says it does not accept funding from frontier AI companies, although those companies provide free model access and tokens. In its stated approach to risk assessment, the nonprofit acknowledges a central limitation: voluntary evaluators still depend on companies to grant access and permit publication of sensitive evidence. It therefore promises to describe constraints, conflicts, evidence quality and material redactions.

METR has previously challenged Anthropic’s reasoning. In a May 2026 review, it concluded that an Anthropic risk report did not adequately support part of its argument about automated research and development. That history shows the evaluator can disagree publicly, but the new embedded model will be tested only when readers can compare METR’s conclusions with a clear account of what it was allowed to examine. For now, the development is a stronger audit mechanism—not an audit result and not proof that Anthropic’s safeguards are effective.