Anthropic lets outside evaluators in to check AI risks

Anthropic has given outside evaluators employee-level access to check risks in its AI models. Six staff members from the non-profit group METR now work inside its offices and can use its hardware and internal systems, with access similar to the company's own risk team.

The idea came from Anthropic's CEO Dario Amodei and won support from Sam Altman and Elon Musk. Anthropic does not pay METR, and the evaluators can publish their reports freely. The company can only ask them to remove things like customer data and model details.

Critics say this setup could lead to conflicts of interest and want stronger government rules. Anthropic also needs to reassure investors that safety issues will not hurt its IPO, which is expected to value the company at over $2 trillion. It is also working with Accenture, with each side putting in at least $1 billion for new safety systems.