“Gambling with Our Lives”: An AI Insider’s Decade-End Warning

Jacob Coxon, a former pretraining researcher at Anthropic (and previously OpenAI), resigned in early September 2026 with a stark public warning: the people building frontier AI “earnestly believe that it could kill us all by the end of the decade.”  He stressed this is “not a marketing stunt,” adding that many executives and senior researchers express far greater fear in private than in public statements.

Coxon’s core concern is not today’s chatbots, but the industry’s rush toward self-improving superintelligence—systems that can recursively enhance their own capabilities.  He warned such systems could soon become “superhuman,” able to hack critical infrastructure, revolutionize fields overnight, and acquire real power and resources without adequate safety guardrails.  In his view, this trajectory amounts to “gambling with our lives,” as companies prioritize speed and competitive advantage over robust alignment and control.

His resignation thread on X gained extra weight when Anthropic’s Alignment Science Lead, Evan Hubinger, publicly affirmed Coxon’s claims. Hubinger wrote: “Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.”  That admission, from inside one of the field’s most safety-focused labs, intensified debate about whether the AI arms race is outpacing our ability to keep powerful systems aligned with human interests.

Coxon called for “pacing agreements” between leading labs to slow the sprint toward autonomous, self-improving models until safety catches up.  He argued that “no other human activity poses this level of danger,” framing the issue as an existential risk that demands industry-wide restraint rather than unilateral acceleration.  His message has become a reference point in the wider “AI doomsday” debate, underscoring how some insiders see a non-trivial chance—above 10%—of human extinction from misaligned superintelligence before 2036.