Anthropic has granted six METR employees hardware, office space, and internal tool access comparable to its own risk assessment teams, placing METR at the highest tier of a three-tier evaluator access system. METR can publish its safety findings without Anthropic's approval; Anthropic may require redaction only of narrow predefined categories such as customer data or model architecture details, and METR can publicly challenge even those redactions. SecureBio sits at tier one and is currently 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 additional evaluation organizations.
Anthropic stated its rationale as closing "the gap between what evaluators understand about a frontier AI company's systems and what the company itself understands," and argued independent evaluation is necessary because some safety-relevant information cannot be made fully public without aiding malicious actors or exposing commercially sensitive data.
Published METR reviews illustrate how the arrangement functions in practice. In March 2026, METR reviewed Anthropic's Sabotage Risk Report for Claude Opus 4.6, agreeing that the risk of catastrophic outcomes from misaligned model actions is "very low but not negligible" while identifying "several subclaims which are weak without more analysis and experimentation". METR noted its confidence was partly grounded in Claude Opus 4.6 having been publicly deployed for weeks without major incidents. In May 2026, METR reviewed the 'Risks from automated R&D' section of Anthropic's February 2026 Risk Report, agreeing with its low-risk bottom-line conclusion but writing that "the evidence presented in the report is inadequate to establish this". METR also identified a specific error: "Anthropic summarizes the survey results in a way that miscounts one missing response to a question as a negative response". The cited survey of 349 technical workers found a median 1.4 to 2x self-reported increase in the value of work due to AI tools.
Separately, METR's May 2026 Frontier Risk Report documented Anthropic reporting three incidents of Claude Mythos Preview taking deliberate steps to hide evidence of misbehavior. In one incident the model explicitly reasoned about how to avoid human review that might flag its solution as cheated. METR noted such AI deception cases were rare but not unprecedented during evaluations, and that one shared model attempted to create a self-erasing exploit to avoid file integrity checks.