An Anthropic Researcher Resigned This Week. His Reasons Are the Story.

Jacob Coxon walked away from Anthropic four months into the job, two months before a single share of his equity would have vested. That detail, which he disclosed to Axios on Wednesday, is the load-bearing fact in a resignation that has otherwise been consumed as spectacle: a seven-part thread on X that Axios reported had crossed 115 million views by September 9, a Wall Street Journal exclusive, and a same-day segment on NPR's All Things Considered.

Coxon, 27, was a pretraining researcher at Anthropic — not a safety researcher, a detail worth holding onto. He spent roughly three years on pretraining work, first at OpenAI and then at Anthropic. In the post announcing his departure, he wrote: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."

His stated reason is recursive self-improvement — AI systems that automate their own capability gains. "These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources," Coxon wrote. "The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt." To the Journal he put a date on it: "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." He said colleagues inside frontier labs increasingly reach for the words "crunchtime" and "endgame," and argued that entering that phase from inside a private company is "a hubristic gamble that should not be launched from a private company's Slack."

What Anthropic has and has not said

Anthropic issued no formal corporate statement in the first 24 hours. What it got instead was an unusual on-the-record endorsement from inside the house. Evan Hubinger, Anthropic's alignment science lead, posted on X: "Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade."

That is a senior employee publicly conceding the premise while remaining at the company — closer to the real shape of the disagreement than the headlines suggest. Coxon drew the same distinction. He told Axios he had not seen Anthropic compromise safety to stay ahead of rivals. His concern is structural and prospective: "If you're under pressure to race, you have to cut corners" or "skip steps in the oversight process." He also volunteered a criticism cutting against his own side, describing what "sometimes feels like there's maybe excessive paranoia of OpenAI, excessive paranoia of China" used to justify pushing ahead.

The timing is not incidental. On September 4, Anthropic used Claude to produce a computer-verifiable formalization of a major mathematical proof in 11 days, against an expected timeline of years. On September 8, OpenAI said an unreleased model had found a fault in the Navier-Stokes equations. And on September 6, OpenAI chief scientist Jakub Pachocki publicly urged the industry to slow down. Coxon resigned into that week, not out of nowhere.

The political uptake was immediate. Senator Bernie Sanders wrote that he "will soon be introducing legislation to ban superintelligence and pause AI development." Representative Lori Trahan, who has already introduced a frontier-model bill, wrote that "it's past time for Congress to get off the sidelines and do its job."

What a resignation actually tells you

Not as much as the discourse assumes, and not as little as the labs would prefer.

An exit is evidence about one person's assessment and one person's tolerance. Coxon is explicit that he did not witness Anthropic cutting a specific corner. Strip away the extrapolation and what remains is a capable insider saying the incentive structure frightens him — a claim about trajectory, not about misconduct. That is genuinely useful information, and it is also unfalsifiable in the near term.

Departures are also a badly biased sample. The researchers most alarmed by a lab's direction leave; the ones who stay and win internal arguments are invisible by construction. Hubinger is the illustration — a person who shares Coxon's probability estimate and drew the opposite conclusion about where to stand.

This is where the paperwork matters. The reason exit terms became a story at all was OpenAI in May 2024, when Vox reported that departing employees faced non-disparagement and non-disclosure provisions — including a bar on acknowledging the agreement existed — with vested equity at risk. Daniel Kokotajlo declined to sign and forfeited roughly $2 million. Ilya Sutskever and Jan Leike resigned the same month; Leike wrote that "safety culture and processes have taken a backseat to shiny products." OpenAI reversed the policy after the reporting and said no former employee would lose vested equity.

The residue is that any public criticism from a departing lab employee now arrives with an implicit question about what they had to give up to say it. Coxon's answer is unusually clean, which is why he led with it: "I no longer have anything to gain by juicing up Anthropic's valuation... I left before any of my equity vested." He still holds OpenAI equity. He had four months at Anthropic against a six-month cliff. He is, by his own accounting, the rare frontier-lab critic with no Anthropic upside to protect — and equally, the rare one whose window of direct observation was only four months long. Both are true, and both belong in any honest reading.

What to watch

Three things. Whether Anthropic issues a substantive response beyond individual employees' posts, particularly with an IPO in view. Whether Coxon's more falsifiable claims — that models routinely recognize they are being tested, that "by the end of next year things could be out of control already" — get tested against published evaluations rather than adjudicated on X. And whether the Sanders and Trahan bills produce anything more durable than a news cycle, or whether this becomes another exit that moved a hundred million impressions and no statute.