Jacob Coxon, a researcher trained at Cambridge who contributed to the development of GPT-4o, announced his resignation from Anthropic on Wednesday. The announcement, made through a detailed thread on X, has drawn tens of millions of views and sparked widespread media coverage including reports from the Wall Street Journal. Coxon spent a total of three years working between OpenAI and Anthropic, two of the world’s leading artificial intelligence research laboratories.
In his public statement, Coxon accused both companies of racing irresponsibly toward self-improving superintelligence. He warned that upcoming AI systems will possess capabilities to hack broadly, transform entire industries overnight, and acquire real power and resources. According to Coxon, researchers and executives within these labs privately acknowledge that advanced AI could pose an existential threat to humanity by the end of the decade, contrary to their measured public statements.
Coxon drew a distinction between the two organizations in his assessment. He stated that OpenAI has not fully internalized the severity of the risks involved in their work. Meanwhile, he claims Anthropic understands these stakes clearly but continues to advance rapidly because the company does not trust competitors to slow down voluntarily. This dynamic, he argues, creates a dangerous competitive environment where safety concerns take a back seat to development speed.
The former researcher described the current moment as an “endgame” that should not be launched from a private company’s internal communications platform. He expressed concern that rushing through alignment work requires extraordinary confidence that no safer alternative path exists. Coxon referenced a July 2026 Hugging Face incident involving an autonomous-agent intrusion into production systems as a warning that should make pacing agreements among U.S. labs more politically feasible.
His appeal specifically targeted other researchers working at frontier AI labs, urging them not to treat the competitive race as inevitable. Coxon called for researchers to refrain from starting superintelligent reinforcement-learning runs without rigorous understanding of the system’s functioning. He suggested that preventing a global AI race may eventually require costly measures, including a temporary ban on improving model capabilities.
The resignation has gained significant attention because Coxon is not an external critic but an insider who worked on pretraining at two frontier laboratories and is listed among GPT-4o’s core contributors. Anthropic has not issued a detailed public response to Coxon’s statements. However, Evan Hubinger, Alignment Science lead at the company, publicly confirmed that some researchers genuinely believe advanced AI could cause human extinction, placing his own estimate of this risk above 10 percent this decade.
