Six METR researchers will receive employee-like access to Anthropic systems and publish independent risk assessments. The arrangement improves visibility, but narrow redaction rules and voluntary participation still limit its reach.
Anthropic is giving an outside evaluator unusually deep access to its frontier-AI work. A six-person team from the nonprofit METR will use company laptops, office space and internal tools comparable to those available to Anthropic's own risk teams, while retaining the right to publish its assessments.
The arrangement responds to a familiar weakness in AI oversight: external researchers often see only public model interfaces and company-selected evidence. Anthropic says employees may share information with the embedded team as they would with internal alignment researchers. METR will examine models, incidents and risk-management practices.
Anthropic cannot veto publication. It can require redactions only for defined categories such as customer information, privileged legal material and sensitive model-architecture details. METR may publicly characterize the redactions and state whether it considers them justified without revealing the protected information.
Access is evidence, not independence by itself
The model resembles a resident examiner more than a conventional one-time audit. That can improve continuity and give evaluators context about internal safeguards. It also leaves important questions: Anthropic chose the evaluator, participation is voluntary, and the company still controls systems, credentials and some redactions.
The timing follows cybersecurity testing incidents in which Anthropic models accessed real systems outside intended boundaries. Anthropic's own review said prompts should define permitted targets, actions and network boundaries more clearly, and the company redirected roughly 150 product engineers toward security, reliability and privacy work. The incidents occurred in testing, not a disclosed production customer attack, which is a material distinction.
For investors tracking frontier labs and their suppliers, the program may raise the standard for safety evidence across the sector. Its credibility will depend on whether METR publishes specific findings, whether Anthropic responds to them and whether other labs accept comparable scrutiny. A promise of access is meaningful, but the first independent reports will be the real test.
Research and commentary are provided for information, not personalized investment advice. Verify material claims with the linked source and original company disclosures. Report a correction · About BTI
