What we know

Anthropic is giving METR researchers employee-level access to its internal systems. This access enables METR to independently evaluate and publish reports on Anthropic’s AI risks, incidents, and operational practices. Anthropic has waived its right to block or censor these publications, allowing METR full editorial freedom in sharing their findings. METR researchers receive employee-level access to Anthropic’s systems. METR will publish independent assessments of Anthropic’s AI risks and practices without Anthropic’s approval or censorship.

Why it matters

Anthropic’s decision to allow deep external scrutiny by METR represents a rare move toward transparency in AI development. By enabling independent evaluations and uncensored reporting, this partnership could set a new standard for accountability in managing the risks of advanced AI models. It also provides a model for how AI companies might collaborate with third-party auditors to build public trust and improve safety oversight.

What is still unknown

Details about the scope and frequency of METR’s assessments remain unclear. The reporting does not specify what languages or AI models will be covered, nor does it include any timeline for when findings will be published. No independent verification of the accuracy or impact of METR’s evaluations has been reported.