Evan Hubinger, who leads alignment research at Anthropic, wrote on X that in his judgment there is a greater than 10% chance that artificial intelligence kills every human being within the next decade. He added that the company he works for is, in his view, doing its best, but does not yet have a plan to solve the problem and is not clearly on track to have one.

Anthropic, which makes the AI assistant Claude, closed a $65 billion round in May 2026 at a $965 billion valuation, and according to the Financial Times is aiming to go public as early as this autumn, at a value some investors put above $2 trillion. Hubinger is the person inside that company tasked with preventing the very outcome he just put a number on.

The temptation, reading news like this, is to file it away as one more apocalyptic warning from an industry that produces them by the truckload. That would be a mistake. So here is some context.

Earlier this week, Jacob Coxon, a researcher who worked on pretraining the company’s models, resigned from Anthropic. According to the Wall Street Journal, he left because he believes the race toward systems capable of improving themselves is moving forward without accountability — not only at his own company. He told the paper he considers it plausible that, by the end of next year, the situation could spiral out of control.

In his farewell posts, Coxon argued that the people building this technology genuinely believe it could kill everyone, that this isn’t marketing, and that executives and senior researchers soften their language in public while privately admitting they’re afraid. No other human activity, he added, carries a comparable risk.

Hubinger quoted that thread on X and replied that Coxon is right, putting his own estimate at more than 10% within the next decade. He wrote that Anthropic “really do earnestly believe AI could kill all humans,” adding that the company is trying its best but does not yet have, and is not clearly on track to have, a plan to solve alignment for superintelligence.

In a follow-up message he narrowed the claim, pointing to Anthropic’s own latest risk report and describing the danger posed by today’s models as low. What he fears is superintelligence that could emerge from recursive self-improvement — systems designing better versions of themselves, in a spiral that Anthropic itself says is accelerating faster than expected. It isn’t about the chatbot someone will use tonight to draft an email; it’s about a capability today’s models don’t have.

Hubinger isn’t an executive who let something slip. He simply repeated, with a slightly different figure, what the company has said since it was founded.

Anthropic was founded in 2021 by a group of former OpenAI executives, including Dario and Daniela Amodei, who believed that race was being run too recklessly. The founding premise has never been hidden: if extremely powerful systems are going to be built regardless, it is better that they be built by people who fear the consequences.

Consistent with that premise, the company’s CEO says things in public that no other CEO says. At the Axios summit in September 2025, asked directly what probability he would assign to a disastrous outcome, Dario Amodei answered 25%, adding that he was focused on the 75% chance things go very well. He has since refined that reasoning on several occasions, stressing that the risk exists regardless of Anthropic — it’s a feature of the technology, not of any single company — and that Anthropic’s very existence is meant to reduce it.

But it is also a position with no precedent in any other regulated industry. A pharmaceutical company that assigned its own drug a double-digit probability of lethal side effects would not sell it — it would disclose that to regulators and pull it from the market. Here, the statement is public, repeated, signed by the top of the company, and produces no regulatory consequence at all. The market listens and keeps buying.

What this means for Anthropic’s IPO

Anthropic’s numbers describe an acceleration with no recent parallel. In February 2026 the company was worth about $380 billion; by May, $965 billion — nearly tripling in under four months. Annualized revenue topped $65 billion by the end of July, roughly seven times the level at the end of 2025, with more than $11.5 billion in the second quarter alone, against $787 million in the same period a year earlier. The May round drew Sequoia, Altimeter, Dragoneer and Greenoaks, but also Blackstone, Fidelity, Baillie Gifford, and the memory-chip suppliers Samsung, SK Hynix and Micron.

If the listing really does arrive this autumn, those statements won’t stay posts on a social network — they will end up, in some form, in the risk-factors section of the prospectus, because a company going public has to disclose the known risks of its business, and the public positions of its top executives are, legally speaking, known information. How to write a risk factor that contemplates the extinction of the species is a problem no legal department has ever had to face before.

Then there’s the regulatory angle. When the U.S. Commerce Department imposed export restrictions on the most advanced models in June 2026, Anthropic had to suspend access to two of its flagship systems for three weeks. Recursive self-improvement itself has so far stayed outside the scope of both European and American rules, which are focused on transparency, data, and high-risk uses. Such an explicit stance from the very people building these systems is exactly the kind of material that shifts lawmakers’ attention.

Finally, it’s worth being clear that there is no scientific consensus on the probability that artificial intelligence causes human extinction, and Hubinger’s estimate is not a forecast — it is a subjective judgment posted on a social network, not the output of a peer-reviewed model. Figures like this have circulated for years in the AI-safety research community, informally called p(doom), and they range from under 1% to more than 50% depending on who states them. What remains true is that this particular figure wasn’t voiced by an outside critic or a competitor, but by the very executive whose job is to prevent that outcome — who, in the same breath, admitted he still doesn’t know how.

Editor’s note

This article was originally published in Italian on money.it by Flavia Provenzani on September 09, 2026 as «“Oltre il 10% di probabilità che l’AI uccida tutti”. Cosa c’è di vero nell’allarme di Anthropic?». It has been translated and adapted for an international audience by the Money.it International desk.