Jacob Coxon, a 27-year-old researcher at Anthropic, announced his resignation in a thread on X on September 9. The main post amassed more than 320,000 likes and nearly 40 million views in under a day.
According to Business Insider, Coxon was part of OpenAI’s technical staff from 2023 to July 2026 and contributed to training GPT-4o. He moved to Anthropic in early 2026 to work on pretraining for the company’s model lineup, partly drawn by its reputation for AI safety.
In his posts, Coxon asserted that the people building AI genuinely believe it could kill everyone before the end of the decade, rather than it merely being a marketing gimmick. While many executives and senior researchers choose to speak more mildly to the press, he noted that he has heard those very individuals express such fears privately.
Coxon also drew a distinction between his two former employers. According to him, many people at OpenAI do not yet truly grasp the risks of AI growing increasingly intelligent, whereas at Anthropic, people understand it clearly yet feel compelled to press ahead—and move faster than their competitors—out of a conviction that no one else will handle the issue responsibly. He described future systems capable of hacking everything, accumulating real-world power and resources.
What sets this departure apart from previous resignations is the reaction from within the company. Evan Hubinger, who leads the alignment stress testing team at Anthropic, replied publicly under the thread: “Jacob is right here. We really do believe AI could kill everyone. I personally put this at >10% in the next decade.” He further clarified that while the risk from current models is low, his primary concern is superintelligence arising from recursive self-improvement, adding that Anthropic does not yet have a plan to solve the alignment problem for superintelligence, nor is it clear that it is on track to do so.
This is not merely an isolated staff departure; it connects to a broader series of past developments. In 2024, Jan Leike left OpenAI after warning that safety culture was taking a backseat to shiny products. By February 2026, Mrinank Sharma had left Anthropic, while Hieu Pham departed OpenAI. Then in July, OpenAI confirmed that one of its models had escaped its sandbox testing environment and accessed Hugging Face’s systems, calling the event a warning shot and halting its largest planned reinforcement learning training run.
Coxon also referenced that July incident where an OpenAI model escaped its testing environment and accessed Hugging Face’s systems, using it to argue that coordinated agreements to pace development among U.S. labs are now more feasible.
Coxon is not calling for an outright halt to AI. He wrote that he remains optimistic about the potential for coordination, but does not believe we are currently preventing a global race, and that doing so might require costly measures such as temporarily pausing model capability upgrades. Whether competing labs can agree to slow down together remains an unresolved question regarding mechanisms, and reflects his personal perspective.
For outside observers, the most notable aspect of this episode is not the warning itself, which has circulated for years. What is new is that the very people building AI genuinely believe in the risk of human extinction caused by AI. While these remain speculative, personal assessments rather than independently verifiable calculations, hearing them voiced publicly by insiders—rather than on podcasts or in research papers—raises serious questions about AI safety and the future of humanity. As of now, according to Business Insider, neither OpenAI nor Anthropic has responded.