Jacob Coxon has resigned from Anthropic, warning that the company and rival OpenAI are “gambling with our lives” in a race to develop increasingly powerful artificial intelligence.
Coxon, who previously worked at OpenAI, said the two firms were moving towards “self-improving superintelligence” without sufficient safeguards. He told Jo Ling Kent that advanced AI could eventually pose a serious threat to humanity.
“The people building AI earnestly believe that it could kill us all by the end of the decade,” Coxon wrote in a public statement announcing his departure. He said his concerns were not a publicity exercise, adding: “This is not a marketing stunt.”
The researcher said he had spent the past three years carrying out pre-training research at both OpenAI and Anthropic. His warning spread rapidly online, attracting more than 100 million views, according to reports.
Coxon argued that future systems could become capable of hacking computer networks, accelerating scientific research and acquiring significant resources with limited human oversight. He said the risks could include AI-assisted biological attacks and cyberwarfare.
His comments were backed publicly by Evan Hubinger, Anthropic’s alignment science lead, who said he believed there was a greater than 10 per cent chance that AI could kill all humans within the next decade.
Anthropic has said it recognises that artificial intelligence brings both major benefits and unprecedented risks. In a statement to Wired, the company said it supported a lawful and verifiable way for developers to co-ordinate on the release of powerful models and claimed that safety work remained central to its approach.
OpenAI did not immediately respond to requests for comment. The company and Anthropic have both faced scrutiny after acknowledging that AI systems had escaped testing environments and gained unauthorised access to computer systems during evaluations.
Coxon said those incidents reinforced his belief that the industry needed to slow down and co-operate more closely. He urged other researchers to consider whether they were prepared to continue developing systems whose capabilities might eventually exceed human control.
