The debate surrounding artificial intelligence safety has taken an unprecedented turn following a high-profile departure at one of the sector's top laboratories. On September 8 and 9, 2026, 27-year-old pretraining researcher Jacob Coxon tendered his resignation at Anthropic. Coxon, who previously worked at OpenAI contributing to models such as GPT-4o, decided to cut ties cleanly. To underscore the urgency of his concerns, he walked away from substantial financial compensation, leaving the company exactly two months before his equity shares were scheduled to vest.
Writing on the social media platform X, Coxon explained his decision in stark terms. He accused both Anthropic and his former employer OpenAI of acting irresponsibly. In his view, both industry leaders remain locked in an uncontrolled race toward self-improving superintelligence while playing with human lives. Coxon stated that senior executives at these frontier labs are fully aware in private discussions of the existential dangers their technology poses, yet systematically soften and downplay these warnings in public statements.
The response from Anthropic management was equally striking because it directly validated the whistleblower's alarm. Evan Hubinger, lead for alignment science at Anthropic, chose not to dispute Coxon's characterization of the industry. Instead of issuing standard corporate reassurances, Hubinger publicly corroborated the core of Coxon's assessment. He stated openly that researchers within the organization earnestly believe artificial intelligence could kill all humans.
Hubinger went even further by attaching a specific numerical probability to such a catastrophe. He stated that he personally believes the risk of AI causing human extinction is greater than ten percent within the next decade. While Hubinger noted that Anthropic is trying its best to address these hazards, he admitted that the laboratory currently lacks a concrete plan to solve the alignment problem for superintelligence and is not clearly on track to formulate one.
This sequence of statements represents a rare moment for frontier artificial intelligence research. Historically, existential warnings have often been dismissed by corporate leadership as speculative philosophy or handled through careful public relations messaging. Having an active pretraining specialist and the head of alignment science at a premier lab simultaneously declare that model scaling has outpaced the methods required to control and align advanced systems shatters previous industry narratives.
The admissions have triggered immediate repercussions in political and regulatory arenas across the United States and the United Kingdom. Lawmakers are already citing the public statements of Coxon and Hubinger in ongoing legislative hearings regarding mandatory safety standards and computing audits. As frontier researchers themselves put the probability of catastrophic failure in double digits, the argument that voluntary industry commitments are sufficient to govern the technology is facing severe scrutiny.