A researcher who spent the past three years working on artificial intelligence at OpenAI and Anthropic has resigned from Anthropic, accusing both companies of recklessly racing toward self-improving artificial intelligence despite concerns about the potential consequences for humanity.
Jacob Coxon announced his resignation on X, saying he had worked in pretraining research at both companies. He accused the two AI firms of failing to act responsibly as they compete to develop increasingly powerful systems.
“Neither company is acting responsibly,” Coxon wrote. “They are racing straight to self-improving superintelligence and gambling with our lives.”
Coxon argued that the AI industry is approaching a point where systems could become capable of hacking computer systems, transforming entire industries and acquiring significant real-world power and resources.
He said the risks are understood by many people working in the field, but argued that competition between leading AI companies is preventing them from slowing down.
“The people building AI earnestly believe that it could kill us all by the end of the decade,” Coxon said, describing the concern as something more serious than a public-relations tactic.
According to Coxon, the situation differs between OpenAI and Anthropic. He argued that many people at OpenAI have not fully internalized what he described as the “civilizational stakes,” while researchers at Anthropic understand the risks but remain caught in a race with competitors.
“At Anthropic, the stakes are well-understood, but they are locked in a race to get there first,” he wrote.
Coxon said companies were effectively operating on the assumption that another laboratory would continue developing increasingly powerful AI if they slowed down, leading them to pursue the technology themselves despite the risks.
He criticized the idea of accelerating the development of superintelligent AI from within private companies, saying that efforts to solve AI alignment should require a much higher level of confidence that the technology can be safely controlled.
Coxon also called for greater coordination between AI companies and governments. He suggested that preventing a global race could eventually require significant measures, including agreements among U.S. AI laboratories to slow development or a temporary ban on improving model capabilities.
His comments have drawn attention from researchers within Anthropic.
Evan Hubinger, an alignment-science lead at Anthropic, publicly responded to Coxon’s warning by agreeing with his assessment that AI could pose an existential threat. Hubinger said he personally estimates there is more than a 10% chance that AI could kill all humans within the next decade.
Hubinger also said Anthropic is trying to address the risks but does not yet have a plan to solve the alignment problem for superintelligent AI and is not clearly on track to do so.
The response is notable because it came from a current Anthropic researcher rather than an outside critic.
Neither company had issued a public corporate response to Coxon’s resignation as of Wednesday. Anthropic and OpenAI did not immediately respond to requests for comment from news organizations covering Coxon’s remarks. TechCrunch reported that Anthropic had not immediately responded to a request for comment, while Fast Company said both companies had been contacted for confirmation of Coxon’s employment and had not responded.
Both companies have previously acknowledged concerns about the pace of AI development. In July, following a statement signed by more than 1,300 employees from frontier AI companies calling for deliberate limits on the pace of AI development, Anthropic said its own research pointed to the need for tools to pace the frontier. OpenAI similarly said AI development could eventually advance quickly enough that the world would need mechanisms to control its pace.
Coxon ended his statement by urging researchers working at AI laboratories to reconsider whether they should continue developing increasingly powerful systems under current conditions.
He questioned whether researchers should simply accept that the technology race is inevitable or instead use the current moment to push for different safety requirements before the development of superintelligent AI advances further.
