Anthropic gives six METR researchers employee-like access for independent safety reviews
The embedded team may inspect internal systems and publish assessments without an Anthropic veto, although defined confidential material can be redacted.
In this article
What employee-like access includes
Anthropic has onboarded six researchers from the nonprofit METR with company hardware, office space and access to internal tools broadly comparable to its own risk teams. Employees are encouraged to share relevant information, except in defined categories such as customer data, legally privileged material and model-architecture information. METR will assess risks, incidents and practices. Anthropic decides which information falls within those protected categories.
The design predates the current agreement
In September, Anthropic chief executive Dario Amodei proposed embedded evaluators with internal tools, publication rights and only narrow redactions for protected information. He argued that outsiders need operational visibility to verify safety commitments rather than relying solely on company-selected disclosures.
The test is what the public eventually learns
METR says its mission is to inform the public about advanced AI capabilities and risks, and that it does not accept payment for frontier-risk assessments. It has previously reviewed Anthropic risk reports and tested internal monitoring systems. METR also says Anthropic and other companies have provided access and tokens for evaluations, research and engineering.
Embedding researchers is more substantial than examining a finished model from outside because it can expose process, context and incidents. Its public value will depend on what this embedded team can publish, what redactions remove and whether its claims can be checked.
