Anthropic gives METR employee-like access to assess model risks

The arrangement puts a third-party evaluator inside Anthropic’s environment, showing one way labs are trying to prove stronger safety checks.

Anthropic has given outside evaluator METR unusually deep access to its internal environment so it can assess the risks of the company’s models, a step that matters because AI labs are under pressure to prove stronger safeguards as systems get more capable. METR, a nonprofit that has also worked with OpenAI, Google, Meta and others, now has staff embedded at Anthropic with access to hardware and internal systems. The arrangement appears to fulfill an earlier Anthropic promise to tighten oversight around model safety. It also reflects one of the clearest actions labs have proposed in response to worries about possible damage to the economy, cybersecurity and society from advanced AI.