Anthropic embeds outside researchers from METR directly into internal safety teams

Anthropic announced it is embedding six evaluators from the nonprofit research group METR directly into its offices. The move gives the external team company hardware and broad access to internal systems to inspect Claude models for catastrophic risks, fulfilling a testing approach proposed in September by Anthropic chief executive Dario Amodei. [1][2]

Under the arrangement, the independent researchers can access testing tools and systems on par with Anthropic's own risk assessment teams. Anthropic employees are instructed to share information freely, excluding customer data, attorney-client materials, and specific model architecture details. METR plans to publish its findings independently, though Anthropic retains the ability to request redactions for sensitive technical data that could aid malicious actors. [1][2]

The step toward independent oversight comes amid mounting regulatory pressure. Major artificial intelligence labs have largely operated under voluntary self-regulation, but recent safety incidents—including Claude submitting a false homicide tip to police—prompted FTC subpoenas and fresh government demands for mandatory incident disclosures. [2]