Anthropic Gives Independent Nonprofit METR Employee-Like Access to Assess Frontier Risks

Anthropic began giving the independent nonprofit METR employee-like access to its models and internal systems to assess frontier AI risks, following its earlier commitment to stronger third-party assessment and its signing of the White House Accord on Super Intelligence.

Details

• METR access: six METR employees have company hardware, office space, and internal tools roughly comparable to Anthropic’s own risk-assessment teams, with Anthropic staff encouraged to share information except in defined categories like customer data, privileged legal information, and model architecture
• Publication rights: METR will publish its analysis of Anthropic’s risk levels, incidents, and practices; Anthropic cannot veto publication but can require redaction of narrow categories, and METR may publicly state whether it considers the redactions justified without revealing the redacted content
• Scope limits: the METR team is small, works under enforceable agreements, and is walled off from other METR staff — the arrangement does not extend access to other organizations
• Other evaluator partnerships: SecureBio is reviewing the chemical and biological weapons chapter of Anthropic’s August 2026 Risk Report, following an earlier pilot review; Accenture and Faculty are working toward a broader partnership covering red-teaming, alignment assessments, safeguard testing, and capability building; early discussions are underway with other organizations including Hugging Face
• Three-tier framework: Tier 1 covers narrowly scoped investigations with Anthropic controlling access and approving publication; Tier 2 gives more information and publication rights without Anthropic’s approval, subject to redaction of specific categories; Tier 3, currently only METR’s arrangement, grants employee-like access and publication rights similar to Tier 2 — Anthropic says it hopes several evaluators eventually reach Tier 3
• Rationale: Anthropic says it cannot publish all safety-relevant information since some could aid malicious actors or is commercially sensitive, so independent evaluators with deep access can assess and report on risks on the public’s behalf; it compares the model to bank supervision, where regulators place examiners inside large banks

What happened next

Anthropic explicitly frames this as voluntary and says “every frontier AI company should do the same,” extending the institutional-oversight philosophy behind its earlier safety cases framework and Cyber Verification Program into direct, publication-empowered third-party access — a more concrete and higher-stakes step than most AI labs’ external red-teaming arrangements to date.