Anthropic gives outside AI watchdog “employee-like” access

Anthropic has granted an external artificial intelligence watchdog "employee-like access" to evaluate the risks of its models. The move follows through on a previous commitment by the Claude maker to increase safeguards against catastrophic failures. The nonprofit evaluator, METR, now has a team embedded inside Anthropic's offices. This team has been given direct access to the company's internal systems and hardware to conduct safety assessments.

METR has previously worked with other major artificial intelligence laboratories, including OpenAI, Google, and Meta. Third-party evaluations represent a primary concrete step proposed by artificial intelligence labs to address growing concerns about advanced systems. Lawmakers and researchers worry that increasingly capable technology could pose severe threats to cybersecurity, the economy, and humanity.

Giving external researchers deep internal access allows for independent risk assessments beyond what internal teams can provide. However, the available reports do not specify the exact duration of the arrangement or which specific models METR is currently testing.