Jacob Coxon spent three years building the kind of AI systems that make headlines. On September 8, 2026, he quit — and instead of a quiet exit, he published a seven-part statement on X accusing his former employer and its biggest rival of gambling with human safety. Within a day, that thread had racked up tens of millions of views, a Wall Street Journal interview had gone live, and a senior Anthropic safety researcher had publicly told the world that Coxon was right.
Here’s what happened, what Coxon actually said, how Anthropic responded, and why the story spread as fast as it did.
Who Is Jacob Coxon?
Jacob Coxon is a 27-year-old AI researcher who previously worked at both OpenAI and Anthropic, according to his Wikipedia entry. His role at Anthropic was pretraining research — the stage of building a large language model where a system learns patterns from enormous amounts of data, before any fine-tuning or safety work happens. Coxon’s combined pretraining work across both companies spanned roughly three years, and OpenAI credits him as a core contributor to GPT-4o, according to reporting on his background.
That background matters for how the story landed. Coxon wasn’t an outside critic or a policy commentator. He’d worked inside the pretraining pipelines at two of the industry’s most prominent labs, which is part of why his departure got read as an insider account rather than an outsider’s speculation.
Why Did Jacob Coxon Resign From Anthropic?
Coxon resigned from his position as an AI researcher at Anthropic on September 8, 2026, criticizing both Anthropic and OpenAI for what he called “gambling with our lives.” He announced the decision himself, in a thread posted on X, rather than through a company statement or a leaked memo.
His opening lines set the tone for everything that followed: he said neither company was acting responsibly, and that both were racing toward self-improving superintelligence. He wasn’t arguing that Anthropic’s safety teams are faking their work — as The Next Web’s account of the thread lays out, his actual claim was narrower and arguably more uncomfortable: that the safety work happening inside these labs isn’t matched to the competitive pressure they’re under, and that researchers can see the risks clearly while continuing anyway because they assume a rival will move faster if they don’t.
Coxon also disclosed a detail that shaped how people read his motives. He said he resigned roughly two months before his Anthropic equity was due to vest, meaning he had nothing to gain financially from the company’s valuation rising after his departure — though, as Axios reported, he still holds equity from his time at OpenAI. He raised that point himself, seemingly to pre-empt the obvious accusation that a disgruntled or profit-seeking ex-employee was just making noise on the way out.
He also framed his exit as a departure from the AI industry altogether, not a move to a competitor or a rival lab.
What Did Jacob Coxon Say About OpenAI and Anthropic?
Coxon didn’t single out one company. He wrote that he’d spent the previous three years doing pretraining research at both OpenAI and Anthropic, and that neither company was acting responsibly. “They are racing straight to self-improving superintelligence and gambling with our lives,” he wrote, in the post Deadline quoted from his thread.
His warning went further than internal caution. He argued that the systems now being built will soon be capable of hacking almost anything, transforming entire fields overnight, and acquiring real-world power and resources. Coxon also pointed to a July incident in which an OpenAI-linked model reportedly breached Hugging Face, the open-source developer platform — describing it, per Quartz’s coverage, as one of the “warning shots” he believes has made coordination between U.S. labs more realistic.
He was careful, though, to separate what he was describing from marketing. “The people building AI earnestly believe that it could kill us all by the end of the decade,” Coxon wrote — adding that this isn’t a marketing stunt, and that the same executives who sound measured in public express fear privately. That’s a claim about what insiders believe, not an independently verified fact about what AI will actually do — a distinction worth holding onto as this story gets retold.
What Does He Mean by Self-Improving Superintelligence?
Coxon’s central worry isn’t today’s chatbots. It’s a hypothetical future capability: AI systems that can improve their own intelligence with limited human direction, rather than being upgraded by human engineers version by version. Self-improvement in this sense means a system advancing itself without meaningful human oversight — a capability that remains out of reach today, even as labs actively work toward more capable and more autonomous systems.
“Superintelligence” is the term researchers use for a system that would exceed human capability broadly, not just at narrow tasks like chess or protein folding. It’s a widely debated theoretical milestone with no agreed benchmark — as CNN noted, some researchers think it’s already partly arrived, others doubt it will ever happen the way it’s usually described. Coxon’s argument is that the industry is accelerating toward that threshold faster than its safety research is maturing, and that once a system can meaningfully improve itself, human control over the pace and direction of AI development gets a lot harder to guarantee.
Why Is He Worried About the AI Race?
Coxon’s underlying point wasn’t that Anthropic’s or OpenAI’s safety teams are dishonest. It was, in the words of one analysis, that the safety work is inadequate to the competitive pressure both companies operate under — and that researchers inside these labs can see the hazards clearly and keep working anyway, because they believe a rival will move faster if they don’t. That’s a structural critique of the industry as much as of any single company: no lab believes it can unilaterally slow down without handing the frontier to a less cautious competitor.
Despite that grim framing, Coxon didn’t argue the situation is hopeless. He said a global AI race remains difficult to prevent, but suggested it could still be slowed through coordination between labs — potentially including a temporary pause on improving model capabilities. He proposed cooperation as a realistic option, not a guaranteed outcome.
What Happened After His Resignation?
The reaction was fast and, by social media standards, enormous. Coxon’s thread reportedly drew nearly 76 million views overnight, a surge amplified by a same-day Wall Street Journal interview with him. That figure later climbed past 100 million views as the story spread, per follow-up reporting.
What made it more than a viral post was what happened next inside Anthropic itself. Evan Hubinger, Anthropic’s Alignment Science Lead, responded publicly on X, writing that Coxon was correct and that “we really do earnestly believe AI could kill all humans,” as Forbes reported. Hubinger put his own estimate of the risk at above 10 percent over the next decade, and added that Anthropic was trying its best but didn’t yet have a plan to solve alignment for superintelligence — and wasn’t clearly on track to do so.
Hubinger followed up shortly after with a clarification aimed at narrowing what he’d meant. He cited Anthropic’s own risk reporting and said the danger posed by models available today is low, and that what actually worries him is superintelligence arising from recursive self-improvement — something he said is happening faster than he expected.
Another Anthropic researcher, Samuel Marks, also publicly backed the substance of Coxon’s warning in the days that followed, according to multiple outlets covering the fallout.
Do Other AI Researchers Share His Concerns?
Coxon’s resignation landed in the middle of an already active debate about AI safety, rather than starting one from scratch. Within a day of his post, Daniel Kokotajlo — a former OpenAI researcher known for his work on AI forecasting — discussed similar concerns during an appearance on Joe Rogan’s podcast, focusing particularly on how competitive pressure between the U.S. and China could push labs toward cutting corners on safety.
Yoshua Bengio, one of the field’s most cited researchers, also weighed in publicly, noting that scientists working directly inside frontier labs often have visibility into emerging risks well before those risks become apparent to outsiders.
More broadly, as Newsweek’s explainer points out, the disagreement in AI safety circles usually isn’t about whether safeguards exist — OpenAI, Anthropic, and other frontier labs all maintain testing regimes, formal risk frameworks, red-teaming programs, and deployment restrictions. The dispute Coxon has reignited is whether those safeguards will still hold up once AI systems become dramatically more capable than what’s available today.
What Does Anthropic Say?
Anthropic hasn’t issued a full corporate rebuttal to Coxon. A company spokesperson offered a general statement instead: “We have always been transparent that AI will bring both enormous benefits and unprecedented risks,” per CNN’s report. Beyond that, the company’s public response has largely come through Hubinger’s posts rather than an official press statement addressing Coxon by name.
It’s also worth noting the backdrop this is playing out against. Coxon’s resignation arrived as Anthropic approaches an IPO that analysts expect to be one of the largest on record, with the company reportedly targeting a valuation in the range of $2 trillion and leaning heavily on its safety-focused reputation to attract investors. That timing is part of why the story got as much attention as it did — a safety-focused resignation lands differently when it coincides with a major capital event for the company involved.
There’s also useful context in Anthropic’s own origin story. The company was founded by former OpenAI employees who left over concerns about that company’s safety practices, arguing they could build AI more responsibly on their own. Coxon’s departure, in other words, echoes the kind of exit that led to Anthropic’s founding in the first place — just aimed at Anthropic this time.
How Serious Are These AI Risks?
This is the part of the story that’s easiest to overstate, so it’s worth being precise about what’s established and what isn’t.
What’s confirmed: Coxon resigned, said what he said, and a senior Anthropic safety researcher publicly agreed with the substance of his warning while offering his own personal probability estimate. That exchange happened, in public, on the record.
What’s not confirmed: any of the numbers being thrown around — Hubinger’s “above 10 percent,” Coxon’s “by the end of the decade” — are individual researchers’ subjective estimates, not measured probabilities or an industry consensus. There’s no instrument that produces an extinction-risk percentage the way there is for, say, a weather forecast. These are informed opinions from people close to the technology, not settled facts.
What’s genuinely uncertain: whether recursive self-improvement is actually close, how fast it might arrive if it does, and whether current alignment techniques would scale to a system that’s meaningfully smarter than its creators. Researchers disagree on all three points, and Coxon’s resignation didn’t resolve that disagreement — it just made it more visible.
What Happens Next?
There’s no clear resolution in sight. Coxon says he’s leaving the AI industry entirely, so his direct influence on Anthropic or OpenAI’s roadmaps ends with his resignation. What continues is the pressure his exit added to an already growing conversation about whether frontier AI development needs external coordination or regulation, rather than relying on individual companies to self-limit.
Anthropic’s IPO process will likely keep this story attached to the company’s public image for a while, given how directly Hubinger’s comments cut against the reassurance investors typically want to hear. And with Kokotajlo, Bengio, and others weighing in within days of Coxon’s post, this looks less like an isolated incident and more like one entry in an ongoing pattern of insiders speaking out — a pattern likely to keep shaping how the public understands what’s actually happening inside frontier AI labs.
Frequently Asked Questions
Why did Jacob Coxon resign from Anthropic?
He said neither Anthropic nor OpenAI, where he’d worked over the previous three years, was acting responsibly given how fast both companies are pushing toward self-improving AI systems. He described the situation as both companies “gambling with our lives.”
Who is Jacob Coxon?
He’s a 27-year-old AI researcher who did pretraining work at both OpenAI, where he’s credited as a core contributor to GPT-4o, and Anthropic, before resigning in September 2026.
What did Jacob Coxon warn about?
He warned that AI labs are racing toward self-improving superintelligence — systems capable of advancing their own capabilities with limited human oversight — without safety work that matches the pace of that race.
Did Jacob Coxon work at OpenAI?
Yes. He worked on pretraining research at OpenAI before moving to Anthropic, and he’s credited among OpenAI’s contributors to GPT-4o.
What is self-improving superintelligence?
It refers to a hypothetical AI system capable of improving its own intelligence largely on its own, and eventually exceeding human capability across most domains. Neither capability currently exists, but Coxon and others argue the industry is moving toward it.
Does Anthropic agree with Coxon’s concerns?
Anthropic’s Alignment Science Lead, Evan Hubinger, publicly agreed with the substance of Coxon’s warning and estimated a greater than 10 percent chance of AI-caused human extinction within a decade, while noting current models pose low risk. Anthropic as a company hasn’t issued a formal rebuttal, only a general statement about AI carrying both benefits and risks.
Is Coxon’s prediction about AI danger a fact or a warning?
It’s a warning — his own assessment and that of some colleagues, not an independently verified or measured outcome. No scientific consensus currently assigns a specific probability to AI causing human extinction.
Why has this resignation gotten so much attention?
Because it wasn’t dismissed or ignored — a senior Anthropic safety researcher corroborated it within a day, the timing coincided with Anthropic’s approach to a major IPO, and the post reached over 100 million views, drawing in other prominent AI researchers within days.
