Anthropic gives six METR researchers an inside view and gives up its publication veto
Anthropic says outside researchers now have company hardware, office space and risk-team-level tools. They can publish assessments, subject to narrowly defined redactions.
Outside evaluators move inside the lab
Anthropic says it has onboarded six employees from METR, the nonprofit that evaluates frontier AI models, with company hardware, office space and access to internal tools and systems roughly comparable to its own risk-assessment teams. The October 15 announcement turns an earlier commitment to outside evaluation into a specific embedded arrangement.
The publication terms are the consequential part. Anthropic says METR will publish its analysis of the company’s risk levels, incidents and practices, and that Anthropic cannot veto publication. The company retains a right to require redactions within narrow, predefined categories, including customer data, attorney-client privileged information and model-architecture information. METR can publicly describe those redactions and say whether it considers them justified, provided it does not reveal the redacted material.
That is broader visibility than a model API alone provides, but it is not unrestricted public access. The engagement covers a small team operating under enforceable agreements and Anthropic’s security standards. Other METR employees are siloed from its work. Anthropic’s footnote says architecture information can sometimes be shared when highly relevant to risk assessment, with explicit internal approval.
Three tiers, with different rights
Anthropic describes three evaluator tiers. The first covers narrowly scoped investigations, with publication requiring company approval. The second adds information access and publication without prior approval, while retaining defined redaction rights. The third adds the employee-like access announced for METR.
SecureBio is currently in the first tier, reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report. Anthropic also describes work toward a broader partnership with Accenture and Faculty, and early discussions with other organizations, including Hugging Face. Those discussions should not be read as equivalent embedded access already granted to every named partner.
The arrangement is voluntary. The sources establish Anthropic’s announced access and publication rules; they do not yet provide METR’s resulting independent assessments or prove that the program will prevent unsafe model behavior. The next substantive evidence will be the findings METR publishes and how any redactions are characterized.