Anthropic gives METR employee-like access for AI risk assessments

Anthropic said on 15 October that it had begun giving a six-person team from METR, a nonprofit that evaluates frontier AI models, employee-like access. This includes company hardware, workspace and internal systems roughly comparable to those used by its risk-assessment teams, with limited exclusions such as customer data and model architectures.

METR will be able to publish analyses of model risk levels, incidents and practices. Anthropic says it cannot veto publication, though it can require redaction of narrowly defined sensitive information. The company also described more limited work with SecureBio and Accenture and a three-tier access framework for evaluators. This is a voluntary oversight arrangement, not a published evaluation result.