A former Anthropic researcher claims, without evidence, that China already has spies inside the big US AI labs

By Zendoric · October 10, 2026.

Jacob Coxon, who spent three years researching model pretraining split between OpenAI and Anthropic, said on the PBD Podcast —the show hosted by entrepreneur Patrick Bet-David— that he is "90% sure" China has already planted spies inside both companies. In his words, "it's almost certain they already have spies there… there still isn't any serious background check for this kind of thing." Coxon did not name anyone or provide any document, incident or evidence to back the accusation.

Coxon left Anthropic in September. In his resignation announcement on X he explained that, after those three years in pretraining research, he believes "neither company is behaving responsibly" and that both are "racing straight toward a self-improving superintelligence, gambling with our lives." On the podcast he went further: he called AI a "superweapon" and warned that society still has not grasped the scale of the issue.

His argument has a recognizable logic: if China can steal US advances at any moment, no temporary lead matters much, and speeding up development to "win the race" would be pointless if the rival can copy the result. By the same reasoning —and here he qualifies that this is only part of China's progress, not all of it— he argues that slowing the American pace would also slow China, because much of its progress would come from copying or stealing American work.

Two things are worth separating. One is the verifiable fact: a former employee with access to sensitive research at two of the world's most important labs has made this accusation in public, on an interview podcast, not in a piece of investigative journalism or an intelligence report. The other is the accusation itself, which neither Coxon nor the Yahoo Finance article backs up with names, documents or specific incidents. The source carries no reaction from Anthropic or OpenAI, nor any outside corroboration of that "90%", which is, at best, a personal estimate.

As industry context, the debate over whether China "copies" or "catches up with" Western AI has spent months mixing verifiable data with self-serving framing: we have already seen how reports on China's military AI muscle risk confusing stated ambition with demonstrated capability. Coxon's accusation fits that same pattern: it is plausible —industrial espionage between technological powers is nothing new— but, on what has been published so far, it is not a proven fact.

What matters, beyond whether specific spies exist or not, is the underlying argument: espionage as the missing variable in the "speed versus safety" calculation that dominates the AI debate in Washington, almost always framed in black and white —accelerate so as not to lose the lead over China, or slow down to reduce risk. If a talent leak or a counterintelligence failure can neutralize any lead in very little time, what really needs governing is not just the pace of releases, but access control over model weights, training data and source code: a matter of corporate and state security, not just benchmarks.

That, for us, is the deeper reading: in the short term, geopolitical competition over AI will keep producing episodes like this one —unproven accusations, real leaks, noisy researcher departures— because the technology has become a national security asset before mature mechanisms exist to govern it as such. That is part of the price of this transition. But managing this phase well requires distinguishing evidence from noise: demanding proof before assuming guilt, auditing access controls inside the labs instead of merely debating the speed of releases, and not losing sight of the fact that neither blind acceleration nor media panic brings the technology closer to its long-term promise: a well-governed AI is the one that can actually deliver the health and abundance at stake.