Anthropic gives METR internal access for AI risk reviews
Illustration, not documentary evidence of the event.
Anthropic is giving six researchers from METR, a nonprofit that evaluates frontier AI models, employee-level access to examine its models, safety practices and risk management. According to techcrunch.com, Anthropic says the researchers have company laptops, office space and internal tools broadly comparable to those used by its own risk assessment teams.
The arrangement also gives METR publication rights. Anthropic says it cannot veto METR’s assessments, although it can require redactions within narrowly defined categories, including customer data, attorney-client privileged material and information about model architectures. METR can publicly describe those redactions and challenge their justification without disclosing the protected information.
For teams deciding whether to deploy Anthropic’s models, the useful distinction is between access to investigate and evidence that a particular use is safe. This announcement describes an evaluation arrangement. It should not, by itself, settle a deployment decision. A practical next step is to look for METR’s published findings, identify which model and behavior each assessment covers, and compare that scope with the tasks and permissions planned for the deployment.
The publication terms matter because “independent review” can describe different levels of scrutiny. The report describes three tiers at Anthropic: narrow reviews requiring publication approval, reviews with more information and no publication approval requirement, and the highest tier, which adds employee-level access. METR is currently the only evaluator in that highest tier. Buyers assessing an evaluation should therefore check both what the reviewer could inspect and whether the company could control publication.
The arrangement could make outside scrutiny more informative by letting researchers examine internal practices alongside model behavior. Its practical value will depend on what the resulting assessments establish. When reading them, check whether conclusions support your intended use, whether material questions remain unresolved, and whether METR describes any redactions as limiting its account. Those are concrete criteria for deciding how much weight to give the review.