Anthropic announced that it has started giving employee-like access to a team from METR, a nonprofit that evaluates frontier AI models. The decision follows the company's earlier commitment to stronger independent, third-party assessments of its models and safety practices, as well as its signing of the White House Accord on Super Intelligence. METR will publish the results of its risk assessments, while Anthropic is also working with SecureBio and Accenture and is in early discussions with organizations including Hugging Face.
The Case for Deeper Third-Party Access
As models have become more powerful, they have required stronger safeguards and more challenging risk assessments. Anthropic says it wants to be publicly accountable but cannot share all safety-relevant information because some of it could be useful to malicious actors or is commercially sensitive. Giving third-party evaluators deep access is intended to close the gap between what external groups understand about frontier AI systems and what the company itself knows.
What Evaluators Need
Anthropic identifies four requirements for effective frontier model evaluation: expertise in assessing model capabilities, independence, enough access to meaningfully evaluate risks, and freedom to publish observations and assessments. Evaluators should not hold equity in the company they are evaluating or be financially or operationally dependent on it. Beyond access to the latest models, they also need visibility into how models were trained and how safeguards function.
Inside the METR Engagement
Anthropic has onboarded six METR employees with company hardware, office space, and internal access comparable to its own risk assessment teams. They can access most information except customer data, attorney-client privileged material, and model architecture details. Anthropic employees are encouraged to share information with them as they would with colleagues working on alignment risk.
Publication Rights and Redaction Limits
METR will publish its analysis of Anthropic's risk levels, incidents, or practices, and Anthropic cannot veto publication. Anthropic can require redaction only for narrow, predefined categories, and METR is free to publicly characterize whether those redactions were justified without revealing the underlying information. The partnership is limited to a small embedded team under enforceable agreements, while other METR employees remain siloed from that team's work.
A Three-Tier Framework
Anthropic is developing three tiers of access for independent evaluators. In the first tier, evaluators conduct narrowly scoped risk investigations with company-defined access and can publish findings only with approval. The second tier allows publication without approval while Anthropic can still require removal of specific categories of information, and the third tier matches the METR arrangement with employee-like access and similar publication rights.
Other Evaluator Partnerships
SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report after completing and publishing a pilot review of the previous risk report. Accenture and Faculty are working with Anthropic toward an ambitious partnership that includes evaluating and red-teaming models, conducting alignment assessments, testing safeguards, and building capabilities across organizations. Anthropic says these partnerships remain in early stages and involve more limited information and publication rights until trust and resources are established.
Toward an Industry Standard
Anthropic compares this approach to bank supervision, where resident examiners work inside large banks to identify public risks in advance. The company is embedding evaluators voluntarily and believes every frontier AI company should do the same. Over time, Anthropic hopes to have several independent evaluators at the third tier, each publishing its own analysis of the company's risk posture and key decisions.
Anthropic's announcement offers a concrete example of how a frontier AI developer can expand external oversight without exposing sensitive information. The partnership with METR tests a model that combines employee-like access with independent publication rights and limited redaction authority. Its long-term impact will likely depend on whether this approach can be scaled across multiple evaluators and adopted by other companies.