Anthropic grants METR internal access for safety audits
Anthropic has granted six staff members from evaluation nonprofit METR full internal access to audit its frontier AI models, establishing a new baseline for independent safety oversight.
In a push for external accountability, Anthropic has onboarded six employees from the AI evaluation nonprofit METR, giving them company hardware, dedicated office space, and internal system access matching its own risk teams. The embedded auditors operate under strict privacy firewalls but hold broad freedom to evaluate models, analyze risk postures, and publish their findings publicly without Anthropic holding veto power over their final reports.
This arrangement represents the third and highest level of a new three-tier evaluation model established by the AI builder. Under Tier 1, external auditors conduct targeted risk reviews subject to publication approval. Tier 2 expands data visibility and allows evaluators to publish independently, subject only to strict predefined redactions. Tier 3 combines this publishing freedom with deep operational access. Alongside METR, Anthropic is partnering with SecureBio to review the chemical and biological weapons chapter of its August Risk Report, working with Accenture and Faculty on red-teaming and alignment assessments, and holding early talks with Hugging Face.
The initiative builds on commitments under the White House Accord on Super Intelligence and draws inspiration from banking regulation, where resident examiners monitor systemic financial risks from within institutions. To prevent leaks, evaluators are restricted from viewing proprietary model architectures or customer data, though Anthropic encourages staff to treat Tier 3 evaluators like internal alignment colleagues when assessing frontier capabilities and potential failure modes such as alignment faking.
For AI safety practitioners and developers, this hybrid supervision model offers a workable template for independent verification without compromising proprietary trade secrets. By establishing pre-agreed rules for redaction and independent publishing, Anthropic demonstrates how frontier labs can open their internal development pipelines to external scrutiny, providing the broader AI ecosystem with verifiable proof of model safety practices.
This is our own summary of reporting by Anthropic
