Jacob Coxon, whose pretraining research took him through OpenAI and then Anthropic, has said he is quitting, citing the way the field is chasing systems that improve themselves. In a thread on X, he said rival labs were “racing straight to self-improving superintelligence and gambling with our lives.” Evan Hubinger, Anthropic’s Alignment Science Lead, replied publicly, and he did not push back on the core of that warning.
That reply is the part of this story worth weighing carefully. Departures over safety objections happen often enough in this industry that any single exit rarely changes the picture. A senior researcher still employed at the lab in question publicly agreeing with the substance of an exiting colleague’s fear is a different order of event.
Posting on X, Hubinger said that he and people he works with “earnestly believe AI could kill all humans,” and gave his own figure as above 10 percent inside ten years. Forbes reproduced both men’s posts on September 9. By his account the company is “trying its best” while still lacking any plan for what he termed “alignment for superintelligence,” and he does not consider Anthropic “clearly on track” to find one.
Read closely, Hubinger endorsed the plausibility of catastrophic risk from future systems, not Coxon’s specific timeline. In a follow-up post that cited the company’s most recent risk assessment, he said the danger posed by models available today is low: “What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought.” Coxon’s own claim went further. He gave his account to the Wall Street Journal, Forbes reports, saying he expects the world is heading toward “a lot of the most aggressive of these scenarios where by the end of next year things could be out of control already.”
Coxon, who has worked inside both labs, also drew a contrast between them. He said OpenAI staff have not “deeply internalized the civilizational stakes,” while at Anthropic the stakes are “well-understood,” yet the company stays “locked in a race to get there first,” on the theory that “no one else will act responsibly, so they must do it themselves, despite the risk.”
That contrast sits uneasily against Anthropic’s own public identity. The company has built its brand around being the safety-focused counterweight to faster-moving competitors, and a resignation grounded in exactly that concern cuts against the pitch. Forbes’s account carries no on-record response from Anthropic to either Coxon’s exit or Hubinger’s remarks. If the company has offered one elsewhere, it was not part of what Forbes reported.
The exit lands alongside a wider push inside the field for a coordinated slowdown. A July statement called “Pacing the Frontier” carried signatures from Anthropic co-founders Dario Amodei and Jared Kaplan, OpenAI Chief Scientist Jakub Pachocki, and Meta AI chief scientist Shengjia Zhao, arguing that industry and government may need room to “buy time” against emerging risks even as competitive pressure pushes against restraint. Pachocki argued along the same lines in a post on Sunday, saying no lab has alignment and monitoring in good enough shape to justify scaling flat out for much longer.
A resignation carries more weight than a survey answer: Coxon gave up a research role at a frontier lab to make this point, a cost an anonymous poll respondent never pays. That is the only reason this story clears the bar for coverage. It is also the limit of what it proves. What changed this week is the size of Coxon’s personal conviction, not the state of the evidence on whether superintelligence is actually near or containable. Anyone weighing which lab to build on should treat Hubinger’s admission, that Anthropic still lacks a plan for aligning superintelligent systems, as a question to keep asking rather than a claim resolved by one social media post.
Forbes, reported by Siladitya Ray, September 9, 2026.
