An Anthropic researcher just resigned in public, and the industry's reaction says more than his warning did

On the evening of September 9, Jacob Coxon posted on X, under his longstanding handle @hilbertspaess, that he had resigned from Anthropic. His post opened plainly: he had spent three years doing pretraining research, first at OpenAI, then at Anthropic, and he did not believe either company was acting responsibly.

The post spread fast, reportedly reaching tens of millions of views within a day, amplified by a Wall Street Journal interview published the same week. Coxon gave up equity that had not yet vested in order to speak on the record rather than staying quiet until his position was secure.

What Coxon actually said

Coxon's argument is not that Anthropic's safety work is fabricated. It is that competitive pressure between AI labs is making that safety work insufficient, regardless of how genuine it is.

In his interview with Axios, Coxon put it directly: being under pressure to race forces labs to cut corners or skip steps in oversight. He also offered a more complicated observation that got less attention than his headline warning, that some of the industry's caution is itself distorted by what he called excessive paranoia about OpenAI and about China, paranoia that can end up justifying pushing capability forward faster rather than slower.

Coxon's central technical concern is recursive self improvement, the idea that once AI systems can meaningfully improve their own training and architecture, progress could compound faster than human researchers can monitor or intervene in it. He has called for coordination among labs and a temporary pause on efforts specifically aimed at improving model capability, as distinct from other AI research.

What Anthropic's own people said back

This is the part worth being precise about, because it is easy to read as either a cover up or a confirmation, and it is neither cleanly.

Evan Hubinger, who leads alignment work at Anthropic, responded publicly rather than through a spokesperson statement, saying Coxon is correct that people inside the company earnestly believe advanced AI could pose a severe risk to humanity, and that he personally would put meaningful odds on a catastrophic outcome this decade. That is a striking thing for an alignment lead at a major AI lab to say on the record, and it happened in response to Coxon, not as a private admission that later leaked.

Coxon has also said, distinctly from his broader warning, that he personally has not seen Anthropic compromise its safety commitments to outlast competitors. That detail is often dropped from summaries of his resignation, but it matters: his argument is about systemic and competitive risk across the industry, not a specific allegation that Anthropic broke a promise internally.

Reporting following the resignation also noted that Anthropic is facing separate questions about its engagement with the UK's AI Security Institute, the government body that evaluates advanced AI systems. That is a live, developing thread and not yet a settled part of the Coxon story itself.

Why this is different from earlier high profile AI safety exits

Coxon's departure has been compared to earlier resignations at OpenAI, including Jan Leike's 2024 departure, where Leike directly alleged that safety work had lost out to product priorities internally. Coxon's framing is different. He is not alleging that either company broke a specific internal commitment. He is arguing that the structure of competition between labs, each individually acting in reasonably good faith on safety, can still produce a bad collective outcome. That is a more structural claim, and a harder one to fact check against any single company's internal conduct.

What remains genuinely unverified

Coxon's most viral line, that AI could kill everyone by the end of the decade, is a prediction, not a documented event. It reflects real, stated concern from people close to frontier model development, including Hubinger's own public odds. It is not a confirmed technical finding, and treating it as one in either direction, dismissing it as hype or repeating it as settled fact, would misrepresent what has actually been established here.

What is established: a credentialed researcher who worked inside two leading labs resigned publicly, gave up compensation to do so, and a senior safety researcher at the company he left corroborated the sincerity of his concern rather than disputing it. What remains open is everything about timelines, likelihood, and what specific policy response would actually address the risk Coxon describes.