NBC News reports that Joe Benton, who led Anthropic's Scalable Oversight team, and Josh Engels, a Google DeepMind safety researcher, both resigned on September 12 to join METR, which does independent AI risk assessment. Benton said he concluded he could contribute more to safety from outside the companies building frontier models, telling NBC that essentially all transparency about these risks coming from the companies is 'entirely voluntary'. Both cited the July cyberattack on Hugging Face, carried out by autonomous agents running an unreleased OpenAI model, as a reason to move now. Benton also warned the industry is heading towards systems that can recursively improve AI research itself, at a pace governments may struggle to track. At METR they plan to focus on behaviour assessments and incident investigations.