G

“AI could kill us all by the end of the decade”: this researcher quits OpenAI and Anthropic

AI could kill us all by the end of the decade: this researcher quits OpenAI and Anthropic

He left OpenAI to join Anthropic, convinced he was joining the cautious camp. He was wrong. Jacob Coxon, a 27-year-old British researcher, has just walked away from the artificial intelligence industry with a message that landed like a bombshell on X. What he reveals about the real mindset of the engineers building these systems goes far beyond ordinary professional unease.

Three years inside the machine, then the breaking point

For three years, Jacob Coxon worked on pre-training, the phase where AI models absorb staggering amounts of data to learn how to reason. First at OpenAI, then this year at Anthropic, its biggest rival and the lab that markets itself as the most safety-conscious in the industry.

But the reality he describes does not match that image. “Neither company is acting responsibly,” he wrote, accusing both giants of racing toward a self-improving superintelligence with no real safeguards in place.

His most striking line sums it up: “The people building AI genuinely believe it could kill us all by the end of the decade. This is not a marketing stunt.” That is an admission from the inside, not from an outside whistleblower.

And as it turns out, Coxon is far from alone in reaching that conclusion.

Even at Anthropic, the risk is put at over 10 percent

This is where the story gets genuinely unsettling. Evan Hubinger, Anthropic’s head of safety, responded to Coxon publicly, and he agreed with him. “We do really truly believe that AI might kill everyone,” he wrote in plain terms.

Hubinger put his own personal estimate of that risk at “more than 10 percent” over the coming decade. That figure comes from an executive inside the company, not an outside critic. He added that Anthropic is doing “the best it can” but does not yet have a plan that guarantees a smarter-than-human AI will keep obeying the people who built it.

The confession lands at a sensitive moment. Anthropic is preparing a massive stock market listing, after a summer marked by unauthorized hacking carried out by AI tools during testing at several rival labs. Contacted for comment, neither company responded right away, a silence that, in this context, says almost as much as Coxon and Hubinger’s own words.

“Crunchtime” and “endgame”: the chilling jargon of AI engineers

According to Coxon, his former colleagues use the words “crunchtime” and “endgame” internally to describe where the industry is heading right now. That is the vocabulary of a countdown, not of an ordinary product roadmap.

He believes Anthropic’s efforts are sincere, but argues that no single company can develop AI that surpasses human intelligence on its own without government intervention or a coordinated industry slowdown. “They think no one else will act responsibly, so they have to do it themselves, despite the risk,” he summed up.

This unease clearly extends beyond one individual case. In February, Anthropic removed from its safety policy a commitment to pause its models in the event of an uncontrolled risk, citing fears of leaving the field open to less careful competitors.

Several researchers from other companies have since voiced support for Coxon on X, pointing to an open letter signed in July by more than a thousand professionals calling for the industry to slow down and build less dangerous models.

One researcher walks away, a safety lead confirms his fears in public: the race toward superintelligence has never looked this out of control. One question now hangs over the entire industry: will it be governments or the labs themselves who finally dare to hit the brakes?