Anthropic gives six METR evaluators employee-like access and says it can't veto what they publish

Anthropic has given six researchers from AI safety nonprofit METR access to its systems roughly on par with its own employees.

In an Oct 15 blog post, Anthropic said the METR team gets company hardware, office space, and internal tools and systems roughly equal to its own risk assessment teams. Apart from narrow categories such as customer data, privileged legal information and model architecture, staff are told to share information as they would with a colleague. Anthropic cannot veto publication of METR's findings and can only require narrow redactions.

Anthropic compared the setup to resident bank examiners. It is also working with SecureBio and Accenture and is in early talks with Hugging Face. AI watcher Andrew Curran wrote that it is "exactly the type of blog post you make just before releasing a shockingly powerful and capable new model," drawing 1,100 likes.