On September 9, 2026, Jacob Coxon, a 27-year-old pretraining researcher who had worked on frontier models at both OpenAI and Anthropic, resigned — and made sure the industry knew why the same day. His warning, sent first to colleagues on Anthropic's internal Slack and then posted as a seven-part thread on X, argued that the labs building today's leading chatbots are "racing straight to self-improving superintelligence and gambling with our lives." Within 24 hours the thread had been viewed well over 100 million times, and Coxon had also given the Wall Street Journal an on-the-record interview about why he left, published within minutes of his own thread going up.
What pushed the story past the usual insider-is-worried news cycle is who backed him up. Evan Hubinger, who leads Anthropic's own alignment science team — the group whose job is to stress-test whether the company's safety techniques actually hold up — replied publicly that Coxon was "correct," and that he personally estimates the odds of AI causing human extinction within the next decade at greater than 10 percent. For a company whose model, Claude, is used by tens of millions of people every day, having its own safety lead put a number like that in public is unusual. It's worth separating what was actually said from what it does — and doesn't — mean for anyone who just wants to know whether their chatbot is fine to keep using this week.
Jacob Coxon, who studied mathematics and did pretraining research at OpenAI from 2023 to mid-2026 (contributing to GPT-4o) and then briefly at Anthropic in 2026, resigned on September 9, 2026. On Slack and in a viral X thread, he said frontier labs are "racing straight to self-improving superintelligence and gambling with our lives," and that unchecked progress risks human extinction. The thread drew well over 100 million views within about a day, and he gave on-the-record interviews to the Wall Street Journal and Axios. Anthropic's alignment science lead, Evan Hubinger, publicly agreed with the substance of the warning, saying he believes there is a greater than 10% chance of AI causing human extinction within the next decade and that Anthropic does "not yet have a plan to solve alignment for superintelligence." Anthropic's official statement said the company has "always been transparent that AI will bring both enormous benefits and unprecedented risks." None of this reflects a discovered flaw in Claude, ChatGPT, or any other shipped chatbot — it's a dispute about where the industry is headed.
Who is Jacob Coxon?
Coxon studied mathematics before joining OpenAI's technical staff in 2023, where he spent roughly three years on pretraining research and was among the contributors to GPT-4o. He moved to Anthropic in 2026 to do similar pretraining work, reportedly drawn — as many hires there are — by the company's safety-focused reputation. He didn't stay long: by his own account to Axios, he resigned after only about four months there, short of the roughly six-month window Anthropic typically requires before equity begins vesting. "I no longer have anything to gain by juicing up Anthropic's valuation," he told Axios. "I left before any of my equity vested." Sources: [Axios](https://www.axios.com/2026/09/09/anthropic-researcher-ai-warning-interview), [Newsweek](https://www.newsweek.com/anthropic-researcher-quits-warns-ai-could-kill-everyone-12418798).
What he actually said
Coxon made essentially the same argument in three separate venues within about 24 hours. Internally, in a Slack message to Anthropic colleagues, he warned that without more caution and cooperation between labs, superintelligent AI carries "a risk of causing human extinction." Publicly, in a seven-part thread on X, he wrote: "I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." In follow-up interviews with the Wall Street Journal and Axios, he sharpened the timeline, telling the Journal: "We're on track for a lot of the most aggressive of these scenarios where by the end of next year things could be out of control already." Sources: [Deadline](https://deadline.com/2026/09/anthropic-jacob-coxon-resignation-artificial-intelligence-1237072134/), [NBC News](https://www.nbcnews.com/tech/tech-news/anthropic-safety-researcher-resigned-warning-rapid-ai-development-gamb-rcna596767).
- Neither OpenAI nor Anthropic, in his assessment, is "acting responsibly" in how it pursues more capable models.
- He believes the industry is on a path toward systems that can improve themselves, faster than safety work is progressing to govern that path.
- He said he personally hasn't seen Anthropic cut safety corners so far, but fears competitive pressure will eventually force it: "if you're under pressure to race, you have to cut corners," he told Axios.
Anthropic's response — and a surprising public endorsement
Anthropic's official line, given to reporters, was measured: a company spokesperson reportedly pointed to Anthropic's history of publicly acknowledging both AI's benefits and its risks, and to the company's ongoing interpretability and risk-reduction research. That's the kind of statement most companies give after an employee resigns publicly and critically. What wasn't typical is what came next: Evan Hubinger, who leads Anthropic's alignment science team, replied directly to Coxon's thread on X rather than staying quiet. "Jacob is correct here — we really do earnestly believe AI could kill all humans!" Hubinger wrote. "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."
Hubinger isn't a random employee reacting online — his team's entire function inside Anthropic is to test whether the company's own safety techniques hold up, which is exactly the kind of internal vantage point that gives a warning like this more weight than an outsider's would carry. It's also notably not a resignation: Hubinger still works at Anthropic, suggesting he sees a greater-than-10%-within-a-decade estimate as a reason to keep doing alignment research there rather than a reason to leave.
The three voices in this story, side by side
| Voice | Role | Core claim | Where it was said |
|---|---|---|---|
| Jacob Coxon | Ex-pretraining researcher, OpenAI (2023–2026) and Anthropic (2026) | Labs are "racing straight to self-improving superintelligence and gambling with our lives"; extinction is a real risk without a change of course | Internal Slack, 7-part X thread, WSJ & Axios interviews (Sept 8–9, 2026) |
| Evan Hubinger | Leads Anthropic's alignment science team | Agrees with Coxon; estimates >10% chance of AI-caused human extinction within a decade; no plan yet for superintelligence-level alignment | Public reply on X (Sept 9, 2026) |
| Anthropic (company) | Corporate spokesperson | Has "always been transparent" about AI's benefits and "unprecedented risks"; points to ongoing interpretability and risk-reduction work | Statement to press (Sept 9, 2026) |
What this actually means if you just use Claude, ChatGPT, or Gemini today
It's worth being precise about what this story is, and isn't. Coxon and Hubinger are both talking about a scenario further down the road — self-improving, superintelligent systems that don't exist yet — not a defect discovered in Claude Opus, Sonnet, or any model currently answering your prompts. Nothing in either of their public statements claims that today's Claude, or any current chatbot, is unsafe for the coding, writing, or research tasks people use it for daily. Anthropic has published a Responsible Scaling Policy since 2023 meant to govern escalating-capability risk as models get more powerful, and Hubinger's stated concern is specifically that the field lacks a proven plan for the superintelligence tier — not that currently deployed models are behaving dangerously.
That said, the disagreement is real, and it's coming from inside the building rather than from outside critics. It's also not without precedent: researchers have left OpenAI, Google DeepMind, and Anthropic before over similar concerns, and Anthropic itself was founded in 2021 by former OpenAI staff who wanted to pursue frontier capability research more cautiously. What's different this time is the combination: a resignation thread that reached well over 100 million views in under a day, followed within hours by a sitting alignment lead publicly attaching a specific probability to existential risk instead of downplaying it.
- This is a disagreement about the pace and governance of frontier AI development, not a reported flaw in any shipped chatbot you can use today.
- Anthropic did not fire or discipline Hubinger for his comments, and he has continued in his role, suggesting the company at minimum tolerates open internal debate about extinction-level risk rather than suppressing it.
- Coverage of this story is still moving quickly as more current and former staff at frontier labs weigh in — treat any single number, including the greater-than-10% figure, as one researcher's personal estimate rather than an industry consensus.
"Greater than 10% chance of extinction" and "AI is safe to use today" are not contradictory statements — they're about different time horizons. The first is about future, far more capable systems that don't exist yet; the second is about the chatbot on your screen right now.
01Who is Jacob Coxon?
A 27-year-old researcher who studied mathematics, worked on pretraining at OpenAI from 2023 to mid-2026 (including contributing to GPT-4o), then briefly joined Anthropic in 2026 before resigning after about four months.
02Did Coxon get fired, or did he quit?
He resigned voluntarily. By his own account to Axios, he left before his Anthropic equity had vested, having been at the company roughly four months — short of its standard vesting timeline.
03Does Anthropic agree with him?
Officially, Anthropic's spokesperson gave a measured statement about being transparent about AI's risks and benefits. Separately, its own alignment science lead, Evan Hubinger, publicly said Coxon was "correct" and estimated greater than a 10% chance of AI causing human extinction within a decade.
04Does this mean Claude is dangerous to use right now?
No. Nothing in Coxon's or Hubinger's public statements claims today's deployed Claude models are unsafe for everyday use. The debate concerns whether the industry has a credible plan for much more capable, potentially self-improving future systems — not the products currently shipping.
05Has anything like this happened before?
Yes, in general terms — safety-related departures and warnings from AI-lab insiders have occurred at OpenAI and elsewhere in past years. What's unusual here is a currently employed safety lead publicly corroborating a departing colleague's extinction-risk claim with a specific probability estimate, rather than issuing a denial.
If stories like this make you want to judge these models for yourself rather than take any single lab's word for it — including Anthropic's — LumiChats lets you compare Claude, ChatGPT, Gemini, Grok, and other leading chatbots side by side, and chat with several of them in one place, so you can form your own view of where the technology actually stands today.
