Anthropic Hands AI Evaluator METR Employee-Level Access: Six Staff Get Internal Systems and Can Publish Without Prior Approval

Anthropic said on Thursday it has granted a team from the nonprofit AI evaluator METR employee-level access to its systems, allowing the group to independently assess the company's advanced models, safety practices and risk management. It is one of the deepest access arrangements an external evaluation organization has received from a frontier AI lab.

Six METR employees have been provided company hardware, office space and access to internal tools and systems broadly comparable to those available to Anthropic's own risk assessment teams, the company said. That goes well beyond the api-based testing arrangements most third-party evaluators receive, where labs control what the evaluators can see and run.

The publication terms are the most consequential part. METR will be able to publish independent assessments of risks, incidents and safety practices without prior approval, subject only to limited redactions covering areas such as customer data, privileged information and certain model architecture details. In other words, Anthropic cannot veto a critical finding before it goes public.

The announcement lands just as scrutiny of lab influence over independent reviews has intensified. A Reuters report this week revealed that OpenAI and Anthropic had considered tighter restrictions for third-party evaluators over concerns that they could leak trade secrets, prompting concern among researchers that the labs building the most capable AI systems also hold too much power over who audits them and how.

When reached for comment on that report, OpenAI said its standard practice is to redact confidential information and that evaluators can disclose when it has made significant redactions. Anthropic reportedly told Reuters it has made clear to existing partners that they can control their own review and publishing process. The METR arrangement now puts operational weight behind that stated position.

Anthropic has also said it plans to publish risk assessment results for its models, working with independent evaluation firms including Accenture and Faculty. Critics have long argued that voluntary, lab-controlled evaluation regimes are structurally inadequate for systems capable of autonomous action, a concern amplified by this year's wave of rogue agent incidents that affected more than a hundred third-party organizations.

For the evaluation ecosystem, the deal sets a template that researchers will now push other labs to match: embedded access, real systems, and publication rights that do not require the subject's sign-off. Whether OpenAI, Google DeepMind and others follow suit, or continue treating external review as a controlled disclosure exercise, may say more about the industry's safety maturity than any benchmark score.