What he said
Jacob Coxon, a researcher who spent about three years working on pretraining at OpenAI and then at Anthropic, announced his resignation on 9 September in a post on X, and said he was leaving the AI industry altogether.
He wrote that he had resigned from Anthropic that day, that neither company is acting responsibly, and that they are racing straight to self-improving superintelligence and gambling with our lives. He said the industry is on track for the most aggressive scenarios, in which things could be out of control by the end of next year, and that accepting the race and entering what he called the endgame is a hubristic gamble that should not be launched from a private company’s Slack.
The Wall Street Journal reported the resignation. Anthropic has not published a response.
The distinction he draws between the two labs
The more interesting part of Coxon’s argument is that he does not treat the two companies as equivalent. At OpenAI, as he describes it, researchers have not fully internalised what is at stake. At Anthropic, they have — leadership understands the risks — but the company is locked in a competitive race and believes no one else will act responsibly, so it must get there first.
That is a sharper criticism than the usual one, because it concedes the safety case rather than disputing it. If a company can understand a risk clearly and still not slow down, the problem is structural rather than epistemic, and internal safety culture is not the lever that fixes it.
What he is asking for
Coxon calls for coordination between labs, and says it should extend to costly actions — explicitly including a temporary ban on improving model capabilities. That is a far stronger ask than the disclosure frameworks and evaluation commitments the sector has converged on, and it has no existing mechanism behind it.
It also lands in a week when the same argument has been made from inside the other lab. OpenAI’s chief scientist Jakub Pachocki wrote last week that no lab has solved alignment well enough to keep scaling at full speed. The difference is that Pachocki said it while continuing to scale, which is close to the point Coxon is making.
How much weight to give it
One researcher’s resignation is not evidence about model capabilities, and a departing employee has no more access to the future than anyone else. What it is evidence about is what people inside these companies believe, and that has a bearing on how seriously to read the safety frameworks they publish.
The category this belongs to is small but not empty: several researchers have left frontier labs publicly over the last two years with variations on the same argument. None has changed a release schedule.
What to watch next
Whether Anthropic responds, and in what register. A company that has built its market position on taking safety seriously has an awkward choice when the criticism comes from someone who worked on its pretraining stack and is not disputing its analysis, only its conduct.
