AI Doom Warning Hits 100M Views in 24 Hours

For AI and machine learning practitioners, this story is not about a disgruntled employee—it's about a pretraining researcher with three years of core modeling experience at OpenAI and Anthropic stating that the alignment problem is unsolved while the scaling race accelerates. Evan Hubinger's >10% extinction estimate from inside Anthropic turns a viral post into an evidence point about the limits of current safety research. The technical question now is whether pretraining toward self-improving systems can be governed before capability outpaces safety.

On the night of Tuesday, September 8, 2026, 27-year-old AI researcher Jacob Coxon posted a resignation thread from Anthropic that became one of the most consequential viral moments in the industry's recent safety debate. Within less than 24 hours the thread had gathered more than 100 million views, and instead of the usual dismissal reserved for AI doomsayers, Coxon's warning was amplified and explicitly endorsed by senior researchers, including his former colleague and current Anthropic alignment science lead Evan Hubinger. Coxon's central claim is stark: he spent the last three years doing pretraining research at OpenAI and Anthropic, and neither company, he says, is acting responsibly. In his words, they are racing straight to self-improving superintelligence and gambling with our lives. The speed and scale of the response suggest the post crystallized a private fear that many in the field share but rarely state this bluntly on the record.

“

Evan Hubinger's >10% extinction estimate from inside Anthropic turns a viral post into an evidence point about the limits of current safety research.

Coxon did not present his exit as a personal grievance or management dispute; he framed it as an existential safety warning. He argued that the people building AI earnestly believe the technology could kill all humans by the end of the decade, and that this is not a marketing stunt. He added that many executives and senior researchers soften their language in public but express the same fears privately, and that no other human activity poses this level of danger. The most striking validation came from Hubinger, whose reply stated: "Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to." That combination of a specific numeric probability and admission that Anthropic lacks a workable alignment plan transformed a single resignation into what many are treating as a public stress test of the safety-first narrative that Anthropic has built since its founding.

Anthropic's corporate response was measured. A spokesperson told media that the company has always been transparent that AI will bring both enormous benefits and unprecedented risks. But the statement did not directly address Coxon's specific allegation that Anthropic lacks a plan to solve alignment for superintelligence or that the company is racing toward self-improving systems. OpenAI, where Coxon previously worked, responded through chief global affairs officer Chris LeHane in a company blog published the day after the post. LeHane called for companies to collaborate and agree to shared standards, and argued there is still time to enact policy that raises the bar for safety. The truncated opening "No first step will be per..." suggests a much longer policy proposal, but the immediate takeaway is that the two leading labs are now being forced to respond publicly to charges that their race dynamic is outpacing safety research.

For the AI research community, the episode matters because Coxon is not an outsider critic or philosopher; he is a pretraining researcher who worked in the core of model development at both OpenAI and Anthropic. Pretraining is the stage where large models absorb capabilities from massive datasets, and it is central to the scaling curve that leads toward more autonomous systems. His claim that the endpoint is self-improving superintelligence, rather than merely more capable assistants, elevates the debate from product safety to governance of systems that could accelerate their own development. Hubinger's >10% estimate is also significant because he is the alignment science lead at Anthropic, a company founded specifically to address alignment risk. If Anthropic's own alignment lead says the probability of human extinction is above 10% and there is no plan yet, that undermines the idea that current safety work has the problem in hand.