ع
Jacob Coxon a 27-year-old researcher who contributed to pre-training models including work on GPT-4o at OpenAI before joining Anthropic earlier this year announced his resignation on the social media platform X on September 9 2026. Coxon wrote that neither company is acting responsibly as they race straight to self-improving superintelligence and gamble with our lives. In the post he emphasised that the people building AI earnestly believe that it could kill us all by the end of the decade adding that this is not a marketing stunt. The researcher who spent three years on pre-training research across both labs described the technology as capable of producing superhuman systems that could hack anything revolutionise fields overnight and acquire real power and resources.
Coxon detailed in his thread that at OpenAI many have not deeply internalised the civilisational stakes while at Anthropic the stakes are well understood but the firm remains locked in a race believing no one else will act responsibly. He had moved to Anthropic from OpenAI in July 2026 viewing it as more cautious yet ultimately concluded that no company can responsibly develop such systems without government intervention or coordinated slowdown. Business Insider reported that Coxon told the Wall Street Journal he does not want to participate in the industrywide rush worried that self-improving systems could spiral out of control and destroy humanity. His departure marks another high-profile exit amid growing unease over the pace of advancement.
Anthropic’s Alignment Science lead Evan Hubinger responded on X confirming that the company really does earnestly believe AI could kill all humans. Hubinger added that he personally assesses the possibility as more than 10 percent within the next decade while noting that despite good intentions the firm lacks a clear plan to avoid uncontrolled superintelligence. The exchange underscores internal recognition of existential risks even as development continues at full speed. Such admissions from senior figures align with Coxon’s account of private fears expressed by executives and researchers who publicly sound more measured.
This resignation follows a pattern of departures from leading AI laboratories including that of Mrinank Sharma who led Anthropic’s safeguards research team and quit in February 2026 warning that the world is in peril from interconnected crises including AI and bioweapons. CNN reported that nearly 1400 AI company employees signed an open letter in July urging the US government to regulate the technology and slow its pace to ensure safety. OpenAI Chief Scientist Jakub Pachocki also cautioned two days before Coxon’s announcement that capabilities are advancing faster than researchers’ ability to monitor and control them. These events highlight persistent tensions between rapid commercial progress and safety considerations.
Pre-training the foundational stage where models absorb vast data sets forms a critical part of developing frontier AI systems at both OpenAI and Anthropic. Coxon’s experience in this area informed his view that the race has become dangerously reckless with insufficient transparency about internal risk assessments. Reports from multiple outlets including Mashable and The Tribune noted that his posts quickly gained attention amplifying calls for greater oversight in the sector. The episode arrives as investors continue to pour capital into AI ventures despite the warnings from those directly involved in building the technology.
Anthropic has positioned itself as a safety-focused alternative to competitors yet Coxon’s critique suggests even its efforts fall short of what he considers responsible development. The researcher’s decision to leave the industry entirely echoes broader fatigue among some technical staff who grapple with the implications of their work. According to industry observers these public exits serve to surface concerns that might otherwise remain confined to internal discussions. Coxon’s thread concluded with advice not to underestimate the power of these emerging systems as they approach superhuman capabilities across multiple domains.
