Anthropic has granted non-profit AI evaluation group METR employee-level internal access and waived its rights to censor external safety reports, setting a new precedent for frontier model oversight. According to Anthropic, the embedded partnership gives six METR staff members access to the company hardware, office space, and the same internal risk assessment tools used by in-house teams. The move directly addresses market skepticism regarding Big Tech internal safety reviews, following reports by Reuters and other outlets alleging that leading developers including Anthropic and OpenAI had previously sought excessive control over independent findings. With Chief Executive Officer Dario Amodei advocating measured development timelines and responding to White House superintelligence commitments, the policy aims to restore public credibility by guaranteeing that external evaluators can publish their findings independently.
The initiative formalizes a three-tier framework for third-party evaluators that dramatically shifts oversight authority away from developer control. Under the baseline Tier 1 model, previously used with partners like SecureBio, Anthropic maintained strict limits on access and required prior approval before reports could be released. Tier 2 permits publications without prior approval while allowing specified redactions. Under the newly created Tier 3 arrangement applied to METR, Anthropic retains no prior veto power over published risk assessments, operational incidents, or organizational practices. The developer can only request redactions for narrowly defined categories, including customer data, attorney-client privileged materials, and model architecture secrets. Crucially, METR preserves the independent authority to publicly disclose whether it considers those redaction requests justified.
Anthropic is also broadening its independent oversight network through red-teaming and alignment partnerships with commercial and open source organizations. The company is pursuing collaborative evaluations with Accenture and Faculty while holding discussions with Hugging Face and other institutions. Over the next six months, this embedded audit model will test whether frontier developers can maintain open scrutiny as standard practice. Competing labs, including OpenAI and Google, face increasing pressure from safety regulators and oversight bodies who may adopt veto-free, resident audits as a prerequisite for commercial deployment. Frontier model developers must now decide whether to risk public exposure from unvarnished risk disclosures or face mounting friction from international policymakers enforcing rigorous third-party auditing rules.