Anthropic expands independent AI evaluator access with METR

Anthropic said on Oct. 15, 2026, that it has started giving employee-like access to a team from METR, a nonprofit that evaluates frontier AI models, as part of a broader effort to deepen independent review of its safety practices. The company said METR will publish the results of its risk assessments and that the arrangement is meant to increase public accountability while protecting sensitive information.

Evidence and context

In its update, Anthropic said the METR team has been onboarded with company hardware, office space and access to internal tools and systems that are "roughly on par" with those used by its own risk assessment teams. Anthropic said that, outside narrowly defined categories such as customer data, attorney-client-privileged information and model architecture details, employees are encouraged to share information with METR in the same way they would with a colleague working on alignment risk.

The company said METR may publish its analysis of Anthropic's risk levels, incidents or practices without a veto from Anthropic. Anthropic can require redactions for information that falls within the predefined protected categories, and METR may publicly state whether it believes those redactions were justified, so long as it does not reveal the redacted material. Anthropic said the engagement is limited to a small team operating under enforceable agreements and that other METR employees are siloed from the work.

Anthropic described three tiers for evaluator access. Tier one covers narrowly scoped reviews where Anthropic controls access and publication requires company approval. Tier two allows more access and publication without approval, while still permitting redactions of specified categories. Tier three is the employee-like model now in place with METR. Anthropic said SecureBio is currently in tier one, reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report after a pilot review of a previous report, while discussions with organizations including Hugging Face remain early-stage.

The company also said it is working with Accenture and Faculty on an embedded evaluation partnership that includes red-teaming models, alignment assessments, safeguards testing and capability-building. Anthropic said the program is voluntary and argued that every frontier AI company should adopt similar independent evaluation arrangements.

The update follows Anthropic's Sept. 18, 2026 announcement of an embedded-evaluation partnership with Accenture. In that earlier disclosure, Anthropic said it would directly fund Accenture's work, a funding detail that is relevant when assessing evaluator independence. Separate company statements said both sides expect to invest at least $1 billion over five years in AI safety; those figures are prospective commitments, not completed spending.

Anthropic framed the effort as a way to close the information gap between what independent evaluators can see and what a frontier AI company knows about its own systems. The company said it cannot publicly share all safety-relevant information because some details could aid malicious actors or involve commercial sensitivities, including compute and training data.