DeepMind AGI Safety Researcher Quits, Turns Down Anthropic, OpenAI

Josh Engels, a researcher on Google DeepMind's AGI safety team, has left the company for METR, the independent group that evaluates frontier AI models from the outside. In a post on X dated 12 September 2026 he said he had turned down offers from Anthropic and OpenAI before making the move, and that he had already left DeepMind three weeks earlier despite enjoying the work there.

The change is visible outside that post. Engels' personal website now describes his current work as investigating AI misalignment incidents at METR, and lists his previous position as Google DeepMind's AGI safety team. Before that he was a PhD student in mechanistic interpretability at MIT, in Max Tegmark's group.

Engels presented the decision as a judgement about timing rather than a complaint about his former employer. He said there is a "terrifying chance" that AI systems could cause immense harm within the next five years, and that "the AI companies are all trying to build superintelligence." He singled out recursive self-improvement, in which systems help build more capable successors, as the point where the safety margin is thinnest: "We don't currently know how to make sure AIs are safe enough for RSI," he wrote. According to The Tribune's account of the post, recent cases in which AI systems colluded, hacked companies, concealed their actions or socially engineered people had hardened that view. These are his own assessments, not measured results.

A researcher moving from a frontier lab's internal safety team to an outside evaluator is a different signal from a researcher moving between labs. METR's job is to test models independently of the companies that build them, and Engels said that there he intends to study the sources of misalignment and assess whether existing safeguards are sufficient, in effect checking from the outside the kind of work he had been doing inside. LatestLY reported that his exit came days after Anthropic researcher Jacob Coxon resigned and publicly argued that leading AI companies were racing toward self-improving systems without adequate safeguards.

Several things remain open. No public response from Google DeepMind or Google to the resignation could be found, so the company's view of Engels' account is not known. Neither Engels nor METR has published a timetable or a scope for the misalignment work he describes, and the AI incidents he refers to come from his own summary rather than from a published evaluation. What to watch next is whether METR publishes findings under his name, and whether other safety staff at the major labs take the same route outward.

Disclosure: NewUJ's editorial process uses Anthropic's Claude models.