Jacob Coxon, a pre-training researcher who spent three years across OpenAI and Anthropic, resigned from Anthropic this week and published a warning that the leading AI labs are behaving recklessly in the race toward superintelligence. “Neither company is acting responsibly,” Coxon wrote on X. “They are racing straight to self-improving superintelligence and gambling with our lives.”
Coxon worked on OpenAI’s technical staff from 2023, contributing to GPT-4o, before moving to Anthropic earlier this year. In his departure note he argued that executives at both labs use measured language in public while privately holding far darker assessments. “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt,” he wrote. He told Fox News the problem is “eminently solvable” and called for urgent global regulation, while putting the risk of catastrophe above 10% within a decade.
Inside Anthropic, agreement rather than rebuttal
What made the resignation land differently from earlier walkouts was the response from Coxon’s own colleagues. Evan Hubinger, who leads Anthropic’s alignment stress testing team, replied 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.” Hubinger added that Anthropic is trying its best but “we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”
Samuel Marks, another safety researcher at the company, backed Coxon’s broader warning while stressing he spoke in a personal capacity. Hubinger drew a distinction between current models, which he assessed as low risk, and future systems capable of recursive self-improvement. Coxon agreed the present threat is minimal but said recursive self-improvement, where AI systems improve themselves, could arrive as soon as next year, with some colleagues estimating six months.
“I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.” – Evan Hubinger, Anthropic
Coxon’s critique of the two labs differed in emphasis. At OpenAI, he wrote, many staff have not deeply internalized the civilizational stakes. At Anthropic, the stakes are understood, but the company is locked in a race to reach superintelligence first, reasoning that no one else will act responsibly so it must, despite the risk. He called accepting that logic and entering the “endgame” a hubristic gamble.
A pattern of departures
The resignation fits a wider pattern of safety researchers leaving major labs. Anthropic amended a key safety pledge this year, dropping a commitment not to train more powerful models without adequate safeguards in place and replacing it with safety roadmaps and risk reports. OpenAI has faced its own internal dissent over the pace of releases, including a sequence of agent-escape incidents earlier this month that pushed the company to add alignment researcher Paul Christiano to its foundation board and safety committee.
Public statements from lab leadership have not settled the question. OpenAI president Greg Brockman described the company’s new GPT-6 Astra model as a generational leap while conceding that AGI has become “more of a mission concept or a spiritual concept” than a testable milestone. Anthropic, meanwhile, has surged to a $965 billion valuation ahead of a planned IPO, with projected annual revenue of $47 billion, pressure that safety advocates say makes voluntary restraint structurally difficult.
The regulatory vacuum
Coxon told Fox News that people inside the industry are “begging” for regulation, a striking position given the labs’ usual resistance. No binding US framework currently governs frontier model training. The government has intervened only selectively, including restrictions on the rollout of Anthropic’s Mythos model earlier this year. In Europe, the AI Act’s frontier-model provisions phase in over several years, and international coordination remains limited to non-binding summits.
The timing is awkward for an industry pitching itself to enterprises and governments. OpenAI claims its Astra model has reached a “critical cybersecurity threshold” where it can independently identify zero-day vulnerabilities and develop attack code without human assistance, a capability that raises the stakes of alignment failures. Anthropic faces separate scrutiny from an American Prospect investigation into a job posting for an enterprise intelligence specialist to track anti-AI activism.
| Figure | Position |
|---|---|
| Jacob Coxon | Resigned; labs are gambling with our lives; risk >10% this decade |
| Evan Hubinger | Agrees; no working plan to solve superintelligence alignment |
| Samuel Marks | Backed the warning, speaking personally |
| Greg Brockman | AGI is now “more of a mission concept” |
What happens next
Coxon said he is leaving the AI industry entirely. His departure does not change model roadmaps, but it adds a public record from someone who worked inside pre-training at both frontier labs and concluded the race cannot be run safely under current incentives. Whether more researchers follow depends less on persuasion than on whether either lab demonstrates, in its next training run, that the safeguards its own staff say are missing actually exist.
The episode also sharpens the contradiction at the center of Anthropic’s pitch: a company founded as the safety-focused alternative to OpenAI now employs researchers who publicly state they do not know how to align the systems they are building. Hubinger’s reply was candid to the point of being an indictment, and it came from the person responsible for stress-testing alignment. Nobody at either lab disputed it.