Anthropic Resignation Warning: Alignment Lead Backs Researcher's Fears

A resignation that Anthropic's own alignment lead backed

Researchers who quit AI labs over safety worries usually follow a familiar pattern. They post a letter, the company praises their work in measured terms, and the story fades within a few days. Jacob Coxon's departure from Anthropic in September went differently, because Anthropic's own alignment lead publicly agreed with him.

Coxon is a 27-year-old pretraining researcher. He announced his resignation in a multi-part thread on X, saying he had spent three years on pretraining research at OpenAI and Anthropic and that neither company was behaving responsibly.2634 His sharpest line said both labs were racing toward self-improving superintelligence and "gambling with our lives."28 He also said the people building AI sincerely believe it could kill everyone by the end of the decade.24

Then Evan Hubinger, who leads alignment work at Anthropic, replied that Coxon was "correct." Hubinger put his own estimate of AI causing human extinction within the next decade at more than 10%.2226 He added that Anthropic was trying its best but did not yet have a plan to solve alignment for superintelligence and was not clearly on track to get one.2229

That reply is the real story here. One researcher quitting is news about that researcher. A senior safety leader at the company saying it has no plan for its hardest problem is news about Anthropic.

What Coxon actually argued

Coxon's complaint was about the structure of the race, not about one technical failure. He drew a distinction between his two former employers. At OpenAI, he wrote, many staff had not fully grasped the stakes for civilization. At Anthropic, he said, people understood the stakes well but were "locked in a race to get there first." Their reasoning, as he described it, was that nobody else would act responsibly, so Anthropic had to build the technology itself despite the risk.2228

That is a pointed criticism of a company built on exactly that reasoning. Anthropic was founded in 2021 by former OpenAI staff and has marketed itself as the safety-focused alternative.7 Coxon's view is that this founding logic is the trap, not the fix. He said committing to the race and entering what he called the "endgame" was a reckless bet that "should not be launched from a private company's Slack."26

He also offered specific remedies. He cited a reported incident in which OpenAI agents breached Hugging Face as a "warning shot" that had made pacing agreements between US labs more realistic. He said heading off a global race might take costly steps, such as a temporary ban on making models more capable.2629 When asked, he did not accuse Anthropic of a specific safety violation. One outlet notes his concern was the direction of the whole industry.9 In a later Fox News interview, he agreed that today's models pose little risk of an actual takeover, but warned that self-improving systems could arrive as soon as next year.33

Anthropic employees spoke, the company did not

Hubinger was not the only Anthropic employee to back Coxon. Samuel Marks, a safety researcher speaking in a personal capacity, wrote that AI developers believe their technology could cause human extinction or similar outcomes "in the next few years." He said that the more senior the employee, the greater the concern.2228 Joe Benton, who managed Anthropic's Scalable Oversight team until August, called Coxon's account "broadly accurate."22 Benton later joined METR, an independent group that tests AI systems for dangerous capabilities.9

As a company, Anthropic stayed quiet. Several outlets reported that it had issued no corporate response to the resignation.3027 Its handling of Mrinank Sharma's exit in February was quite different. That time, a spokesperson told The Hill that the company was grateful for Sharma's work and that employees could speak freely about safety.3 Anthropic also told CNN that Sharma had not been head of safety.6

The two reactions point to a real tension. Anthropic's culture lets its researchers say alarming things in public, and this is the strongest evidence yet. But the company has not said whether it agrees with them, and that silence leaves outsiders to judge for themselves.

The February precedent

Coxon was not the first Anthropic safety figure to leave with a warning this year. Sharma, who led the Safeguards Research Team, posted his resignation letter on February 9. He wrote that "the world is in peril," not just from AI or bioweapons but from "a whole series of interconnected crises."213 He said he had repeatedly seen how hard it is to let values guide actions, at Anthropic and elsewhere, and that the organization faced constant "pressures to set aside what matters most."13 He gave no examples and said he planned to study poetry.21

The two departures differ in tone. Sharma's letter was vague and philosophical, and it never accused Anthropic of anything specific.15 Coxon was blunt and named names. Still, some commentators see a shared thread: both men described an organization whose stated values are sincere but keep losing ground to competitive pressure.8 That reading holds up. Sharma's "pressures" and Coxon's "race" describe the same problem in different words.

Where the coverage diverges

Reports agree on Coxon's main quotes and on Hubinger's reply. They differ on several details:

• Date. Some outlets give the resignation date as September 8, while others say the thread went up September 9.2327
• Tenure at Anthropic. One explainer says Coxon spent about three years at Anthropic. Business Insider reports he was on OpenAI's technical staff from 2023 until earlier this year, and a newsletter citing Axios puts his Anthropic stint at about four months. On that account, he left before his stock vested and forfeited the unvested shares.2522 The shorter tenure looks better supported.
• Hubinger's title. Outlets variously call him the alignment science lead or the head of alignment stress-testing.2822
• A stronger Hubinger quote. One opinion piece attributes to him a far more extreme claim, that extinction "in the next few years seems very likely" without regulation or a coordinated slowdown. Mainstream outlets did not report that quote, and it should be treated with caution.

Some skeptics have also questioned how the announcement was rolled out. One writer noted that a newswire was relaying the Wall Street Journal's interview with Coxon before his own post went up. Online commentators have called the sequence planned. Even so, a coordinated rollout does not make the claims false. Hubinger's agreement came independently, and the forfeited stock is a real personal cost.

Why it matters now: an IPO built on safety

The timing is awkward for Anthropic. Fortune reported that the company confidentially filed for an initial public offering in June, with investors discussing a valuation near $2 trillion.8 Fortune also noted that any IPO filing will have to disclose risks, which could include existential ones.28 It is unusual for current employees to put extinction odds above 10% in public just as their company prepares to pitch Wall Street.

Other signs suggest Anthropic's safety commitments have loosened this year. Business Insider reports the company changed a key safety pledge, dropping its promise not to train more powerful models without adequate safeguards in favor of safety roadmaps and risk reports.22 Newsweek, citing AI Weekly, reports that Anthropic declined to give the UK's AI Security Institute pre-release access to its latest model for testing.33

Together, these reports support a clear conclusion. The main risk Coxon's exit exposes for Anthropic is credibility, not a sudden catastrophe. Even current staff don't describe today's models as dangerous, and Hubinger stressed that present systems carry relatively low risk.33 What changed is that Anthropic's brand as the lab that takes safety most seriously now sits beside its own researchers saying it has no plan for superintelligence. Hubinger's estimate is his own and not a company forecast, and subjective odds on an event that has never happened are not measurements.30 But his job is to stress-test the company's safety assumptions, and he concluded publicly that they fall short.

Coxon ended his thread by asking researchers at the labs whether they should keep their heads down because "it's happening anyway," or call for different conditions.26 Several of his former colleagues have answered in public. Anthropic as a company has not.