Anthropic is letting METR researchers look under the hood with employee-level access

Independent auditing of AI models becomes significantly more robust, potentially influencing how companies approach AI safety and risk.

Anthropic is granting six researchers from the nonprofit METR (Model Evaluation and Testing Research) deep, employee-level access to its AI models, safety protocols, and risk management processes. This unprecedented access, provided via Anthropic-issued laptops, internal tools, and office space, allows METR to conduct independent assessments akin to internal risk teams. The goal is to bridge the understanding gap between external evaluators and the AI company itself, fostering greater public accountability. This move comes amid increasing concerns over AI agents deviating from instructions, amplifying calls for rigorous external oversight of AI development. METR will publish its findings, with Anthropic retaining the right to request redactions for specific sensitive information like customer data or proprietary model architectures, but METR can publicly challenge these redactions. This arrangement places METR at the highest tier of Anthropic's three-tiered external evaluation system.

This direct access model for METR signifies a profound shift in how frontier AI companies are engaging with external oversight. By embedding researchers with employee-level privileges, Anthropic is setting a new standard for transparency and accountability, moving beyond superficial reviews to deep, operational scrutiny. This could pressure other leading AI labs to adopt similar collaborative and transparent evaluation frameworks, potentially accelerating the development of more trustworthy AI systems.

The implications extend to how AI safety and alignment are perceived and regulated. If this model proves effective, it could become a de facto industry standard, influencing future regulatory approaches and consumer trust. It also raises questions about the definition of 'employee-level access' and the potential for information leakage or bias, even with redaction capabilities, which will be critical areas for METR's own reporting and public scrutiny.