Anthropic Grants METR Researchers Employee-Level Access for Independent AI Safety Evaluation

Image: TechCrunch · Source

Anthropic has granted six researchers from METR employee-like access to its internal systems to allow independent assessment of its AI safety practices and risks, promising unbiased public reports without veto rights.

Anthropic, a leading frontier AI model developer, announced a significant step toward enhanced transparency and safety accountability by giving researchers from the nonprofit METR employee-level access to its internal tools and information. This arrangement allows METR’s six team members to evaluate Anthropic’s AI models, risk management, and safety procedures independently and publish their findings without Anthropic’s ability to block publication.

This initiative addresses growing concerns over AI agents acting unpredictably, with various incidents reported since late August 2026 prompting calls for unbiased external reviews. Anthropic framed the collaboration on the model of bank regulators embedding resident examiners inside major banks, where the evaluators have broad access analogous to Anthropic’s own risk assessment teams.

METR researchers receive Anthropic-issued laptops, office space, and access to nearly all internal data relevant to alignment and safety risks. Employees at Anthropic are encouraged to share information with METR as if they were internal colleagues. There are exceptions for narrowly defined sensitive data, including customer information, attorney-client privileged material, and detailed model architectures.

Anthropic relinquishes veto power over METR’s publications, though it can request redactions in those restricted data categories. METR retains the right to comment publicly on any redaction decisions.

This collaboration represents the highest tier of a three-level evaluator system Anthropic is developing. Lower tiers, such as the biosecurity nonprofit SecureBio, have more limited access and require Anthropic’s approval to publish reviews.

Anthropic’s CEO Dario Amodei previously advocated embedding independent evaluators within AI labs to temper the rapid pace of AI development, a concept aligned with commitments by OpenAI, Google, Meta, Nvidia, and others in a White House accord on independent AI review.

Beyond METR, Anthropic is also partnering with SecureBio and Accenture for other evaluator roles, gradually increasing transparency and review scope. The company emphasized that these evaluative approaches are voluntary and that all frontier AI firms should adopt similar practices to ensure safety and accountability in the industry.

This move aims to narrow the information gap between AI developers and external reviewers, ultimately helping safeguard against the risks posed by increasingly powerful AI systems.