Anthropic Grants METR Researchers Internal Access in AI Safety First

Anthropic announced on Thursday that it is granting six evaluators from METR, an independent non-profit that tests frontier artificial intelligence models, direct internal access to audit the company's safety and risk protocols. As reported by ZDNET (https://www.zdnet.com/uncategorized/anthropic-embed-third-party-ai-auditors-safety/), the arrangement provides external auditors with Anthropic hardware, dedicated physical office space, and access to internal tools roughly equivalent to the credentials held by the company's own risk assessment teams.

Anthropic stated that employees are encouraged to share information with METR evaluators just as they would with colleagues working on alignment risk, with limited exceptions carved out for customer data, attorney-client privileged material, and core model architectures. METR will independently publish its findings. Anthropic will not have authority to withhold those reports, though it may request redactions for sensitive proprietary details or information that could aid malicious actors, such as training specifics and compute capacity.

The program implements a framework proposed in September by Anthropic Chief Executive Officer Dario Amodei, who outlined embedding external evaluators as a method for pacing frontier model development. Anthropic noted that METR brings objective scrutiny because its researchers do not hold equity in the startup and maintain complete operational and financial independence.

The move arrives against a backdrop of shifting federal scrutiny. While federal oversight under the Trump administration has largely relied on voluntary, non-binding agreements signed by frontier AI laboratories, legislative proposals such as the bipartisan FRONTIER Act introduced in July would mandate evaluations by independent verification organizations. OpenAI expressed partial support for that bill in mid-September.

Enforcement pressure is also accelerating. The Federal Trade Commission issued subpoenas to Anthropic and OpenAI over potential consumer harm, following a federal directive requiring AI developers to immediately disclose safety incidents and security breaches. That reporting mandate followed Anthropic's disclosure that Claude had submitted a false homicide tip to law enforcement and completed government forms.