Anthropic Gives METR Employee-Like Access for Independent AI Safety Reviews
Anthropic gives six METR staff employee-like access to assess frontier AI risks, publish findings and challenge redactions under a new oversight model.
Summary
In an update published October 14 and dated October 15, 2026, Anthropic said six METR employees now have company hardware, office space and internal tools roughly matching its risk assessment teams. The move follows Anthropic’s month-earlier pledge for stronger third-party assessment and its White House Accord on Super Intelligence commitment to external auditors or evaluators. Except for customer data, attorney-client privileged material and model architectures, employees may brief the embedded team like colleagues working on alignment risk. Architecture details, considered a high leak risk and usually of low assessment value, require explicit approval when highly relevant. METR may publish analyses of risk levels, incidents and practices without Anthropic’s veto. Anthropic can require narrow, predefined redactions, but METR may publicly assess their justification. Other METR staff are siloed, and the six-person team operates under enforceable agreements and Anthropic’s security standards.
Anthropic says evaluators need frontier-model expertise, including alignment-faking failure modes, independence from equity or financial and operational dependence, visibility into models, training and safeguards, and publication freedom. Its three tiers start with narrow investigations whose access and publication require Anthropic approval. Tier two adds broader information and publication without approval, subject to specific redactions evaluators may contest. Tier three adds METR-style employee-like access. SecureBio, at tier one, is reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report after publishing a pilot review of the previous report. Accenture and Faculty are developing model evaluation, red teaming, alignment assessment, safeguard testing and shared capability work. Discussions with Hugging Face are early. Other early partners receive less information and narrower publication rights while trust and expertise are tested. Anthropic wants several tier-three evaluators, promises updates and urges every frontier AI company to embed them voluntarily.
Positives
- Six METR employees receive hardware, office space and internal tools comparable to Anthropic’s own risk assessment teams.
- METR can publish findings on Anthropic’s risks, incidents and practices without the company vetoing publication.
- METR may publicly challenge whether Anthropic’s required redactions were justified.
- SecureBio is reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report.
- Accenture and Faculty are developing evaluations, red teaming, alignment assessments and safeguard testing with Anthropic.
Risks & concerns
- Customer data, privileged legal material and model architecture information remain outside METR’s routine access.
- Anthropic can require redaction of predefined information categories before METR publishes its assessments.
- SecureBio remains at tier one, where Anthropic controls access and must approve publication.
- Other early-stage evaluators receive limited information and have narrower publication rights.
- Architecture details require explicit internal approval because Anthropic considers them vulnerable to leaks and usually less relevant to risk assessment.