A researcher who helped train frontier models at two of the world’s leading AI laboratories has left Anthropic, arguing that the industry’s sprint toward systems that can improve themselves is an unacceptable gamble with human survival.
Jacob Coxon, 27, spent three years on pretraining work—the process of feeding models enormous datasets to raise raw capability—first at OpenAI and then at Anthropic.
In comments first reported by The Wall Street Journal and expanded in a public thread, he said he could no longer take part in what he described as an industry-wide race toward self-improving superintelligence.
He argued that neither of his former employers is behaving responsibly and that the competitive dynamic is putting lives at risk.
Coxon warned that the systems now being built will soon exceed human performance across many domains.
He said they could breach digital defenses, accelerate scientific and industrial change almost overnight, and accumulate real-world resources and influence.
Progress, he added, shows no sign of slowing.
He told the Journal that some of the most aggressive timelines already look plausible, and that events could slip beyond effective human control by the end of next year.
Inside labs, he said, language such as “crunch time” and “endgame” has become common.
He was equally blunt about private attitudes among builders.
People constructing these systems, he claimed, genuinely believe advanced AI could wipe out humanity before 2030.
Public comments from executives and senior researchers are often measured, he said, but the same people express sharper fear in private.
No other human project, in his view, carries comparable danger. He rejected the idea that such warnings are publicity.
Coxon drew a distinction between the two companies.
At OpenAI, he said, many staff have not fully absorbed the civilizational stakes.
At Anthropic, those stakes are widely understood, yet the company remains locked in a race because leaders believe rivals will not act with restraint and therefore feel compelled to reach the frontier first despite the risk.
Accepting that race and treating the next phase as an “endgame,” he argued, is a reckless bet that should not be decided inside a private firm.
His departure drew an unusual public reply from Evan Hubinger, Anthropic’s alignment science lead.
Hubinger said Coxon was right that many at the company earnestly believe AI could kill all humans.
Hubinger put his own estimate above 10 percent within the next decade.
He added that Anthropic is trying hard but does not yet have a plan for aligning superintelligence and is not clearly on track to produce one. Current models, he stressed, present relatively low risk; the concern is recursive self-improvement arriving faster than expected.
The episode adds to a pattern of internal unease at frontier labs.
Anthropic’s CEO has long spoken about catastrophic risk, and some industry figures have called for coordinated pacing of capability growth.
Critics of the resignation notes say competitive pressure from other companies and countries makes unilateral slowdowns difficult.
Supporters argue that if even safety-focused labs concede they lack a solution for superintelligent alignment, the public should treat the timeline as an urgent policy problem rather than a distant science-fiction scenario.Coxon has left the industry.