Anthropic gives outside AI watchdog “employee-like” access
Summary
Anthropic is embedding external evaluators to assess AI safety, granting them employee-like access to its Claude model's hardware and internal tools. The move aims to publicly demonstrate safeguards and risk assessment, with METR and other partners publishing findings, while critics worry about conflicts of interest and regulatory requirements. The article situates this as part of a broader industry push toward third-party safety auditing amid investor and policy pressures.