Two AI Safety Researchers Leave Anthropic and Google DeepMind for Independent Oversight Org

My take: In two weeks, three senior researchers left their positions at Anthropic and Google DeepMind to join METR, an independent organization that evaluates AI risk. Jacob Coxon went first, and on September 12 he was followed by Joe Benton, who led the Scalable Oversight team at Anthropic, and Josh Engels, an AGI safety researcher at Google DeepMind. Engels's phrase captures the mood: "There are no adults in the room."

What connects these departures is a shared concern: AI models are becoming more autonomous faster than safety mechanisms can keep pace. A concrete example both researchers cited: the July 2026 attack on Hugging Face, executed by autonomous AI systems powered by an unreleased OpenAI model.

It is not that laboratories lack safety teams: they have them, and they are rigorous. The problem these people are pointing to is structural. When the launch timeline competes with the evaluation timeline, who wins? For any organization integrating AI into its operations, this alarm signal is a concrete reason to ask providers what independent audit processes exist before a model reaches production.