Anthropic has opened its internal operations to six researchers from the nonprofit METR team, providing employee-level access to enable thorough independent assessments of its AI models' safety and risk management protocols.
- METR researchers receive company laptops, tools, and office space to assess Anthropic AI.
- Anthropic cannot block METR’s publications but can request limited redactions.
- This initiative expands on CEO Dario Amodei’s call for embedded independent reviews.
What happened
Anthropic has granted six researchers from the nonprofit METR team employee-level access to its internal systems. This includes company-issued laptops, office space, and the same tools accessible to Anthropic’s own risk assessment teams. As a result, METR can directly investigate Anthropic’s frontier AI models, risk management approaches, and safety practices.
The researchers operate under conditions similar to internal employees and can freely communicate with Anthropic staff about alignment risks. METR will independently publish assessments of Anthropic’s incidents and practices, with the company relinquishing the right to veto publications outside narrow exceptions such as customer data and privileged information.
Why it matters
This unprecedented collaboration demonstrates a major shift toward openness in the frontier AI space amid escalating concerns over the unpredictable behavior of AI agents. Providing employee-level access to external reviewers closes the information gap often found between internal teams and outside evaluators, which has historically limited independent accountability.
Anthropic’s approach is comparable to regulatory resident examiners embedded within large financial institutions, representing a new model for AI oversight. This transparency model can strengthen public trust by subjecting AI safety claims to rigorous and impartial scrutiny, a crucial step as frontier AI models grow more powerful and complex.
What to watch next
The progress and findings from METR's independent assessments will be closely watched by AI developers, regulators, and the broader public. How METR reports on Anthropic’s risks and incidents, including their use of redactions, will set important precedents for future AI oversight collaborations and transparency standards.
Additionally, other leading AI companies like OpenAI, Google, and Meta, which have committed to similar independent reviews under a White House accord, may adopt comparable embedded evaluation mechanisms. The emerging tiered evaluation system Anthropic is building—ranging from narrow external reviews to METR's full internal access—may become a blueprint for AI risk governance frameworks worldwide.