External AI Watchdog Gets Embedded Access at Anthropic

Anthropic has granted METR, a nonprofit evaluator that audits systems across major AI labs, resident status inside the company with direct access to hardware and internal systems. The arrangement allows for hands-on risk assessment of Claude’s capabilities and safety mechanisms while keeping sensitive technical details confidential.

The move responds to mounting concerns that advanced AI systems could destabilize the economy, infrastructure, and society. It marks the first concrete implementation of a proposal that attracted rare alignment among rival AI executives—Altman, Musk, and Amodei all backed embedded external oversight in September.

Anthropologic positioned the arrangement as a middle ground: proving accountability to the public without exposing commercially sensitive or strategically valuable safety work. The company draws a parallel to bank regulation, where government examiners work on-site at major financial institutions. Skeptics worry the approach creates conflicts of interest and argue stronger government oversight is necessary.