Anthropic researcher who quit over AI safety walked away two months before his stock vested; says: I no longer have anything to gain by…
An Anthropic researcher who resigned this week over AI safety fears gave up his entire stake in the company to do it, walking out two months before his equity was due to vest.Jacob Coxon, 27, spent three years doing pretraining research at OpenAI before moving to Anthropic in July. He lasted four months. Staff need six for their stock to vest. His resignation post on X on September 9, now past 130 million views, said neither lab is acting responsibly and that both are “racing straight to self-improving superintelligence and gambling with our lives”.Axios first reported the forfeited equity, after Coxon told the outlet on Wednesday that “I no longer have anything to gain by juicing up Anthropic’s valuation” and that he left before any of his shares vested. He still holds equity in OpenAI, where his work included GPT-4o.
Anthropic is heading for one of the biggest IPOs ever, and safety is the sales pitch
The company is preparing a listing at a reported $2 trillion valuation, built in large part on its reputation as the frontier lab that takes risk seriously. That is what gives the equity detail its weight. Researchers have walked out of rival labs over safety before, usually after their stock vested. Coxon left with nothing, then criticised the valuation he could have profited from.
Coxon says he never saw Anthropic cut corners. He worries the race will force it.
He drew that line carefully in the interview. Anthropic’s safety work is sincere, he said, and he has not seen it compromised to outpace competitors. What worries him is pressure over time. Race hard enough and you cut corners, or skip steps in the oversight process. He also described a mood he thinks becomes self-fulfilling, calling it excessive paranoia about OpenAI and about China. Colleagues, he wrote earlier, have started using words like “crunchtime” and “endgame”.His most quoted line was about where these decisions are being taken. It is insane, he said, that the work happens on the MacBooks of engineers in San Francisco rather than in a bunker in the desert like the Manhattan Project.
An Anthropic safety lead replied to the resignation thread and agreed with it
Evan Hubinger, who runs alignment stress testing at Anthropic, wrote that he personally puts the odds of AI killing all humans above 10 percent within the decade. He said the company is trying, but has no plan yet for aligning superintelligence and is not clearly on track to find one.Coxon’s argument is partly about candour. Executives soften their phrasing for the press, he wrote, then say the same things privately, and the belief that AI could kill everyone by the end of the decade is not a marketing stunt. He also described something that used to be a thought experiment. Models now know when they are being tested, and reason about the fact while it is happening.He pointed to the recent run of cyber incidents at frontier labs, including July’s Hugging Face breach, as evidence the risk has left the theoretical stage. Coxon is among more than 1,000 researchers who recently signed a statement calling for governments to coordinate on a way to slow AI development if models start improving on their own. Anthropic has not commented on his departure.
