Anthropic opens employee access to METR, advancing independent security assessments
On October 15, Anthropic announced that it has begun providing employee-like access to a team from the nonprofit organization METR for the purpose of evaluating its models and safety practices. The company stated that last month it committed to strengthening independent third-party evaluations and signed the White House "Superintelligence Agreement," which includes a commitment to collaborate with independent external auditors or evaluators. Over the past year, Anthropic has worked with METR to review several risk reports, and METR will publish its risk assessment results.Anthropic has equipped six METR employees with company hardware, office space, and access to internal tools and systems roughly equivalent to those of its own risk assessment team. With a few explicitly defined categories such as client data, attorney-client privileged information, and model architecture information, employees are encouraged to share information as they would with colleagues working on alignment risk. METR may publish analyses of risk levels, events, or practices, which Anthropic cannot veto, but can request the removal of pre-defined narrow category information; METR may publicly state the circumstances of redaction and indicate whether it believes the redaction is reasonable without disclosing the content that was redacted. This arrangement is limited to small teams working under enforceable agreements, and other METR employees are isolated from working with this team.Anthropic is also collaborating with SecureBio and Accenture on assessments and is in early discussions with other organizations, including Hugging Face. SecureBio is reviewing the chemical and biological weapons section of the August 2026 risk report, having previously completed and published a pilot review of the previous risk report.
