Anthropic has embedded six employees from the nonprofit evaluator METR inside the company, giving them hardware, office space and access to internal tools and systems broadly comparable to its own risk-assessment teams. METR can publish its findings without Anthropic’s approval.
The arrangement is meant to let outside specialists inspect powerful models more closely than public testing permits. Anthropic argues that some safety information cannot be released openly because it could help malicious actors or expose competitive details. It compares the setup to resident bank examiners, who receive privileged access so they can identify risks before they threaten the public.
METR has evaluated systems from several leading AI labs and previously reviewed model logs and company equipment while investigating an AI-related compromise of Hugging Face. Its presence at Anthropic is an early implementation of a broader proposal to place external evaluators inside frontier labs. The model still has limits: critics warn that evaluators working so closely with a company may face conflicts of interest, and private assessment is not a substitute for government rules. Its credibility will depend on whether METR can independently choose what to investigate and publish meaningful conclusions when its findings are unfavorable.