Anthropic has onboarded six METR researchers with internal access and publication rights to independently assess its models, incidents, and safety practices.
Anthropic said on October 15 that it has begun giving six researchers from the nonprofit METR employee-like access to examine the company’s AI systems and risk practices. The arrangement gives the evaluators company hardware, office space, and access to internal tools and systems broadly comparable to those used by Anthropic’s own risk-assessment teams.
The access is not unlimited. Anthropic says the evaluators are excluded from certain predefined information, including customer data, attorney-client privileged material, and model architecture details. Employees are otherwise encouraged to share information with the METR team as they would with colleagues working on alignment risk. The company says the METR group is covered by enforceable agreements and is separated from other METR employees, including those working with Anthropic on other projects.
A central feature is the publication arrangement: METR will be able to publish its assessments of Anthropic’s risks, incidents, and practices without the company’s approval. Anthropic may request redactions for specified categories of sensitive information, but METR can publicly describe whether it considers those redactions justified, provided it does not disclose the protected material.
Anthropic places the METR arrangement at the highest of three access tiers it is developing for external evaluators. At the lower tiers, evaluations are more narrowly scoped or publication can require company approval. The company says it is also working with SecureBio and Accenture and is in early discussions with other organizations, including Hugging Face, though those relationships currently have more limited access or publication rights.
This is a voluntary evaluation program, not government oversight. Its significance will depend on what METR can examine in practice and what its published assessments reveal. The arrangement also leaves important limits in place: reviewers cannot freely inspect every category of information, and the company retains a role in requesting narrow redactions. Anthropic says it hopes to embed more independent evaluators as these partnerships develop.