The collaboration allows METR to conduct rigorous testing on Anthropic’s systems, though the arrangement retains specific boundaries. While the safety organization is tasked with publishing findings regarding risk levels and operational incidents, Anthropic maintains the right to mandate redactions. This oversight covers sensitive proprietary elements, including the core architecture of its models and private customer data. By granting these evaluators employee-like access, the company aims to bridge the gap between internal safety protocols and the need for independent third-party verification in an increasingly high-stakes industry.
Anthropic Grants METR Safety Evaluators Internal System Access
Six researchers from the AI safety organization METR are receiving unprecedented access to Anthropic’s internal hardware and development tools. This arrangement mirrors the privileges held by Anthropic’s own risk assessment team, marking a shift toward externalizing the scrutiny of complex model behaviors and potential security vulnerabilities before public deployment.
Comments (0)
No comments yet. Be the first!