Anthropic gives outside evaluators deeper access to its AI safety work

Anthropic said on October 15 that six employees from the nonprofit METR had begun working with extensive access to its internal systems. The arrangement lets outside evaluators examine the Claude developer’s safety practices more closely.

The evaluators have office space, company equipment and access comparable to Anthropic’s internal risk assessors. Anthropic cannot block publication of their conclusions, but it can require removal of narrowly defined confidential information. METR may publicly challenge whether those removals were justified.

The company said outside scrutiny must reach information that cannot safely be released publicly. The initiative follows its September commitment to strengthen independent assessment and earlier work with METR reviewing risk reports.

Anthropic plans to develop additional evaluator relationships through three levels of access. SecureBio is reviewing chemical and biological weapons risks, while work with Accenture and Faculty includes testing safeguards. The company said it hopes more evaluators will eventually receive access comparable to METR’s.