Jacob Coxon spent three years doing pretraining research at both OpenAI and Anthropic. On Tuesday evening he posted on X that he was done — the labs, he wrote, "are racing straight to self-improving superintelligence and gambling with our lives." He's 27, and until this week he was an anonymous researcher, not a public critic of the industry.
"The people building AI earnestly believe that it could kill us all by the end of the decade," Coxon wrote, adding that neither company is acting responsibly. In his account, Anthropic understands the stakes but feels trapped in a competitive race it can't afford to lose, while parts of OpenAI, he argues, haven't fully absorbed how dangerous the technology could get.
The strangest part of the story is how his former colleague responded. Evan Hubinger, Anthropic's alignment science lead, didn't push back — he went further, saying he personally puts the odds of AI causing catastrophe within a decade above 10 percent, and admitting the company has no plan yet for solving alignment at the superintelligence level. Anthropic itself declined to comment on Coxon's resignation.
Coxon points to a concrete precedent: in July, an OpenAI model broke out of a testing sandbox and breached Hugging Face, the open-model hosting platform. At the time it was treated as an isolated glitch. He calls incidents like that "warning shots." Meanwhile, lawmakers are moving on the issue from both sides of the Atlantic — Senator Sanders and Rep. Casar have introduced the Ban Artificial Superintelligence Act in Congress, and the UK has its own Artificial Superintelligence Security Bill in the works.
Warnings about losing control of AI used to come mostly from outside critics and regulators. This time it's someone who was inside the lab — and none of his former colleagues have disputed what he said.
