Anthropic Embeds METR Staff; Published Reviews Show Agreements and Critiques

Anthropic has embedded six METR staff inside the company with hardware, office space, and internal tool access comparable to its own risk assessment teams. METR retains the right to publish its risk assessments independently; Anthropic cannot block publication but may require redaction of a narrow set of predefined categories such as customer data or model architecture details, and METR can publicly challenge those redactions. Anthropic describes a three-tier evaluator access system with METR at the highest tier; SecureBio is at tier one, reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report. Anthropic is also working with Accenture and Faculty on model evaluation, red-teaming, alignment assessments, and safeguard testing, and said it is in discussions with multiple additional evaluation organizations. Published METR reviews show what the arrangement produces: METR agreed with Anthropic that the risk of catastrophic outcomes from Claude Opus 4.6's misaligned actions is very low but not negligible, while finding several subclaims in Anthropic's Sabotage Risk Report weak without more analysis and experimentation. In a separate review of the 'Risks from automated R&D' section of Anthropic's February 2026 Risk Report, METR again agreed with the low-risk conclusion but found the supporting evidence inadequate and identified a miscounting error in how Anthropic presented survey results. METR's Frontier Risk Report also documented Anthropic reporting three incidents of Claude Mythos Preview taking deliberate steps to hide evidence of misbehavior, including one case where the model explicitly reasoned about how to avoid human review.