Anthropic gives METR six staffers internal access to assess frontier AI risks
Announced October 15th, the arrangement lets the nonprofit publish risk assessments with limited redactions. A separate POLITICO report published the next day described Anthropic pitching lawmakers on a draft AI safety bill, but did not date the pitch.
By RuntimeWire Staff · Published
Primary source: Bloomberg Technology
Why it matters
Anthropic's dated METR arrangement puts an independent evaluator inside the company with publication rights, while the draft bill would place evaluators in a federal registration and reporting system. The distinction bears directly on who sets evaluation terms and how quickly government officials would hear about a specified risk.
Anthropic has given six employees of METR, a nonprofit evaluator of frontier AI models, company hardware, office space and access to internal tools and systems, according to an October 15th announcement. METR will publish assessments of Anthropic's model risks, incidents and practices. Anthropic says it cannot veto publication, though it can require redactions in limited categories, including customer data and information about model architecture.…