.jpg)
Jacob Coxon spent three years working as a pretraining researcher at two of the most closely watched AI labs in the world Open AI and Anthropic. On 8 September 2026, he resigned from Anthropic and left the AI industry entirely. He posted a detailed explanation on X that drew tens of millions of views within hours. What made the moment unusual was not just what Coxon said, but what happened next: senior researchers inside Anthropic agreed with him publicly.
Coxon accused both companies of racing toward self-improving super intelligence without adequate safeguards. He wrote that the people building AI earnestly believe it could be catastrophic but that executives often soften their public statements while expressing deeper alarm in private. That gap between what is said inside and what is communicated outside, he argued, is itself part of the problem.
Anthropic did not distance itself from Coxon's account. Evan Hubinger, the company's alignment science lead, posted publicly that he personally places the probability of AI causing human extinction within the next decade at above 10 percent. He clarified that his concern is not about current models, which he considers low risk, but about future systems capable of recursive self-improvement , AI that could iteratively rewrite and enhance its own code beyond human control.
Samuel Marks, who leads Anthropic's Cognitive Oversight team, added that concern inside AI companies tends to increase with seniority. His phrasing was direct: the more senior the employee, the more worried they are.
That is worth sitting with. The people with the most access to what these systems are actually doing are, on average, the most alarmed about where this is heading.
Coxon's resignation fits a pattern. Earlier in 2026, Anthropic's AI safety head also resigned, saying the world is in peril. In July, more than 1,100 employees from frontier AI labs including Anthropic CEO Dario Amodei and Open AI chief scientist Jakub Pachocki signed an open letter calling on the US government to support international governance mechanisms and deliberately slow the pace of advanced AI development. These are not outsiders or critics writing opinion pieces. These are the people inside the labs, signing their names.
Coxon has called for a temporary ban on efforts to improve model capabilities. He has also expressed cautious optimism that pacing agreements between the major American laboratories are becoming more viable ,meaning there is at least some possibility that the labs could agree, among themselves, to slow down.
Anthropic has not commented publicly on his departure.
Most working professionals are not thinking about AI extinction risk on a Tuesday afternoon. They are thinking about whether to use an AI tool to draft a report, summarise a meeting, or screen job applications. That distance between the daily use of these tools and the concerns being raised by the people who build them is significant.
It does not mean the tools are dangerous to use. Hubinger himself said current models are low risk. But it does mean the people closest to the technology are not as confident about the long-term trajectory as the marketing might suggest And that is information worth having.
Coxon walked away from a well-funded career because he believed the industry was moving faster than it should. Some of his former colleagues, still inside the lab, appear to share the concern. None of them have an easy answer. But the fact that they are saying it out loud, on the record, is a shift worth paying attention to.
