Anthropic Update on Independent Evaluators

Anthropic announces it has begun giving a METR evaluation team employee-like access to internal systems, hardware, and office space so those evaluators can perform deep, third-party risk assessments of its frontier models and publish their findings. Six METR staff are embedded under enforceable agreements; Anthropic cannot veto publications but can require redactions for narrowly defined categories such as customer data, privileged communications, and certain architecture details. METR may publicly note and assess whether redactions were justified. The arrangement is intended to close the information gap between an AI company’s internal safety understanding and independent evaluators’ perspectives while protecting commercially sensitive or potentially dangerous details.

Anthropic is rolling out a three-tier evaluator framework: first-tier engagements are narrowly scoped and require company approval to publish; second-tier teams can publish without approval though limited redactions are permitted; the third tier - exemplified by METR - receives employee-like access and broad publication rights. SecureBio is currently in tier one reviewing biothreat sections, and Anthropic is developing deeper partnerships with Accenture and Faculty for red‑teaming, alignment assessments, and safeguards testing, with early discussions with others including Hugging Face. The company emphasizes precautions against inadvertent leaks (especially of architectural data), plans to expand embedded evaluators, and argues that voluntary, embedded independent evaluation should become standard practice for frontier AI firms.