Anthropic has granted six METR researchers employee-like access to evaluate its AI risks and safety practices, expanding independent oversight while retaining restrictions on certain confidential information. (Source: Image by RR)

New Oversight Arrangement Follows Amodei’s Call for Frontier AI Safeguards

Anthropic has begun giving outside AI safety researchers direct access to its internal systems as part of an effort to strengthen independent oversight of increasingly powerful artificial intelligence models. The company announced that six employees from the nonprofit research organization METR have been onboarded with company hardware, office space and access to internal tools comparable to those available to Anthropic’s own risk assessment teams. The arrangement, as noted in an article at anthropic.com, follows CEO Dario Amodei’s recent proposal for stronger external scrutiny of frontier AI development and Anthropic’s commitment under the White House Accord on Super Intelligence to work with independent evaluators.

The embedded researchers will examine Anthropic’s model risks, safety practices and potentially concerning incidents while receiving access to information generally unavailable to outside observers. Anthropic says qualified evaluators need more than access to finished AI models; they also require insight into training practices, safeguards and internal decision-making to meaningfully assess risk. METR will be permitted to publish its findings without Anthropic’s approval, although the company can require redactions involving predefined categories such as customer information, privileged legal material and certain proprietary technical details. Anthropic retains authority to determine which information falls within those protected categories, while METR can publicly describe limitations on the access it received.

Anthropic is also developing a three-tier system for working with external evaluators. The first tier allows narrowly defined investigations with company-approved publication, while the second provides broader access and the ability to publish independently, subject to specific redactions. The third tier, currently represented by METR’s embedded team, provides employee-like access alongside independent publication rights. Additional organizations are participating in more limited arrangements, including SecureBio, which is reviewing chemical and biological weapons risks, and Accenture and Faculty, which are working toward broader evaluations involving model safeguards, alignment and adversarial testing. Anthropic has also held preliminary discussions with other organizations, including Hugging Face.

The initiative represents a move toward continuous external scrutiny rather than relying exclusively on internal evaluations or public safety reports prepared by AI developers themselves. Anthropic compares the arrangement to bank supervision, where independent examiners receive access to confidential information that cannot be fully disclosed publicly. However, the program remains voluntary, and questions about information restrictions, evaluator independence and consistency across AI companies remain unresolved. Anthropic says it intends to expand embedded evaluation partnerships and argues that other frontier AI developers should adopt similar practices as model capabilities and associated risks continue to grow.

read more at anthropic.com