Anthropic gives METR access to audit models and safety practices

Anthropic said it had given six staff from METR company hardware, office space and access to internal tools so they could review its models and safety practices. The company said that arrangement left them roughly on a par with its own risk teams.

Anthropic said customer data, information covered by attorney-client privilege and details of model architecture were not available to METR. METR said it would publish its findings on its own.

Anthropic said it could ask for redactions in predefined categories, but METR could say publicly whether it thought those requests were justified. The company also said not all safety-relevant information could be made public because it might help attackers or because it was commercially sensitive.

Anthropic said this review went further than what it had offered SecureBio, Accenture and Faculty, which it said had less access and tighter publication limits. It also said it had started early talks with Hugging Face.

Anthropic said the move followed its signing of the White House Accord on Super Intelligence, which included a commitment to work with outside reviewers.