Anthropic Grants METR Employee-Level Internal Access for AI Safety Research

Release Overview

Anthropic has announced that it is granting researchers at METR (Model Evaluation and Threat Research), a non-profit organization specializing in AI safety evaluation, "employee-level" access to the internals of its AI models. This partnership provides an environment where independent third-party researchers can closely investigate model behaviors and internal mechanisms.

Objectives and Implementation of the Partnership

The core of this announcement is that METR researchers are now able to conduct safety evaluations of AI models through an environment typically accessible only to internal engineers. By disclosing the model's structure in detail to an external specialized institution alongside internal evaluations, Anthropic aims to drastically improve the quality of AI model safety and risk assessment.

Initiatives for AI Safety

AI transparency and safety are critical challenges in current generative AI development. Anthropic granting METR access to its model internals serves as a significant milestone toward ensuring transparency in safety standards across the entire industry. Through this strong collaborative relationship with METR, the formulation of objective and rigorous safety standards is expected to accelerate.

Future Outlook

With this newly granted access, METR researchers will be able to conduct more practical and in-depth safety evaluations. Anthropic plans to continue deepening its dialogue with independent specialized organizations and research groups, pursuing a path of responsible AI development that prioritizes safety above all else.