Former Anthropic and OpenAI researcher Jacob Coxon said he was 90 per cent certain that China already had spies inside both US artificial intelligence (AI) companies, without providing evidence.

Coxon made the allegation in an interview with entrepreneur Patrick Bet-David on the PBD Podcast, arguing that China could steal advances from leading US AI labs and quickly close any technological gap.

“Right now, if we just stopped, they would try and catch up. But I think one really important thing people don’t realise is that China already has spies at OpenAI and Anthropic,” Coxon said.

“China has the ability to hack at any time. If we pull ahead, that doesn’t give us a big lead. They can just step in, steal our progress at any time and catch up.”

Asked by Bet-David how certain he was that Chinese spies were already inside the companies, Coxon replied: “Ninety per cent. Yeah, it’s almost certain that they have.”

Reacting to Coxon’s comments on social media, Paul Triolo, an expert on China’s technology policy and cybersecurity, said: “Sorry, but this seems really irresponsible.”Coxon’s remarks could also raise concerns inside US AI labs, which rely heavily on talent from China.

Researchers from China accounted for 38 per cent of the top AI talent working in the US, slightly ahead of Americans at 37 per cent, according to the MacroPolo think tank. Three years earlier, Chinese researchers represented 27 per cent of top AI talent in the US, compared with 31 per cent for Americans.

The comments come days after Washington and Beijing agreed to launch an AI dialogue focused on the technology’s “risks and benefits” during Chinese President Xi Jinping’s state visit to the US, with the next round scheduled for November. The two sides also agreed to establish a “bilateral communication channel” to discuss AI-related incidents.

Coxon, who spent three years working on pre-training research at OpenAI and Anthropic, resigned from Anthropic in September, warning that the two companies were racing towards self-improving AI without adequately addressing the risks.

“The people building AI earnestly believe that it could kill us all by the end of the decade,” he wrote on social media in September.

His resignation and comments drew widespread attention and sparked debate over whether leading AI labs were adequately addressing safety concerns, along with calls to slow development.

However, the Trump administration has rejected calls to slow down, citing the global competition with China for AI dominance.

US President Donald Trump has instead backed industry self-regulation, meeting leading AI executives at the White House last month and endorsing a voluntary agreement for companies to monitor and audit their own AI systems. He described the agreement as “morally binding”.

As Chinese companies have closed the gap with US firms in frontier AI models, American technology companies and Trump administration officials have increasingly accused Beijing of copying or extracting capabilities from US models to accelerate its own development.

“The technical word for stealing and copying American AI models is distillation. So, the Chinese distil our models, and they can never get ahead of us,” US Treasury Secretary Scott Bessent said last month.

“It’s kind of like if you’re looking over someone’s shoulder, copying their homework, you can never get a higher grade than they do.”Model distillation is a widely used AI training technique in which a smaller model learns from the output of larger ones. China has rejected US allegations, terming distillation “a common practice in the AI sector”.