An OpenAI engineer is accused of hiding a recording device in the USB port of a television in one of the company's executive conference rooms. According to a Justice Department announcement on October 13, 2026, OpenAI found the device on September 30 and assessed that it had been installed around September 28. The FBI arrested the engineer, Abdulrahman Alfozan, a 35-year-old Saudi national, on October 9.
The complaint, as the department summarizes it, describes earlier questions about meeting rooms and calendars, device-related prompts typed into a chatbot account associated with his spouse, a return to the conference room on October 5, and a second, similar device found in his bag. The charge is using or attempting to use a device to intercept oral communications. It is not an espionage charge, and no court has found him guilty.
On Moonshots with Peter Diamandis, recorded October 20 and released the next day, the host and his regular panel spent a segment on the case with guest Chamath Palihapitiya. Peter Diamandis, founder of XPRIZE and Singularity University, said the story "reads like a spy novel" and stressed that "these are still unproven allegations." He relayed OpenAI's position that it had found no evidence of unauthorized access to model weights, source code or customer data. Model weights are the billions of numbers learned during training; in practice they are the model itself. Diamandis also said new court filings described frequent WhatsApp contact between the engineer and a member of the Saudi royal family who encouraged him to come home and join a Saudi AI initiative.
Then he put a question to the panel: what is more valuable to steal these days, model weights or what is said in the conference room?
A strange target and a bigger contest
Alexander Wissner-Gross, a computer scientist and founder of Reified, found the choice of target "a little bit strange." The usual way to siphon what a model knows, he said, is distillation: training a cheaper model on a stronger model's outputs, including the step-by-step reasoning traces it produces. He described distillation attacks from China as the still-favored route.
Setting the allegations aside, Wissner-Gross saw a "great game" between great powers. In the 1950s it was fought over nuclear secrets, in the 1970s over oil, and in the 2020s over superintelligence. Picking up an earlier thread in the episode about Western and Eastern spheres of influence, he predicted that within a few years, if not already, there will be "a de facto second Cold War between the West and the East over superintelligence." Countries in between, in the Middle East and Europe, would feel pulled both ways, and some would try to stay non-aligned. He expected the pull to be very hard to resist, and said he would expect to see "many stories like this."
"Assume that most of everything we do is being exfiltrated"
Diamandis asked whether frontier labs, the handful of companies building the most capable models, should now be secured like defense contractors. Palihapitiya, founder and CEO of Social Capital and co-host of the All-In Podcast, said this has been happening in Silicon Valley for a long time.
He recalled telling colleagues at Facebook, once the company passed 3,000 employees, "We have spies from at least four countries." People asked what he was talking about. His reasoning was statistical: look at how many staff were on H-1B work visas and which countries they came from, and the odds say some of them are working for a foreign government. His posture then, he said, was that "you should just assume that most of everything we do is being exfiltrated."
His second example came from a company he had invested in whose work was export-controlled. It was hacked and "everything was exfiltrated, like everything." When third-party consultants, the FBI and the NSA came in, Palihapitiya asked how common this was. Their answer, as he described it, was that it happens everywhere. The OpenAI case, he said, is reported to sound like "cloak and dagger, 007 type stuff," but he believes it is happening "literally everywhere in every tech company that matters. The labs obviously matter." His advice was to understand that a very small share of foreign staff are spies and "proceed accordingly."
Dave Blundin, founder and general partner of Link Ventures, said the stakes have changed. If Facebook had spies back then, he said, the stakes now are "a thousand, 10,000, a million times higher," and pressure from foreign governments to place spies will grow with them. He doubted California was built for that. Nuclear secrets are locked down, but California's economy runs on people leaving big companies to start small ones, which he called the "chain reaction" that keeps it going. Treating AI like a nuclear secret, he said, would mean you'd "have to rethink the entire structure."
Diamandis worried about a different reaction. He said he hoped the case would not become "a national meme" that leads the U.S. to turn away foreign nationals. He has long argued that a PhD or master's degree earned in the United States should come with a green card, and said that sending people back would be a problem for the country.
"This time is different"
Wissner-Gross saw one detail as new. OpenAI, he said, used its own AI to work out how the allegedly planted device worked. Historically, he argued, secrets could not defend themselves: nuclear secrets couldn't, and neither could oil. Superintelligence, "in some sense, is the first critical and emerging technology that's potentially capable of defending itself territorially." He called OpenAI's use of its models "breadcrumbs, baby steps" toward an AI that prevents itself from being exfiltrated or ending up in undesirable hands.
Why one overheard sentence could matter
Blundin pointed to the flip side. He cited Meta's Llama 3 and Llama 4 models, which he called "crappy models" but whose research papers were published openly. The changes that worked and didn't work, he said, "you could summarize in like 10 sentences."
He offered a hypothetical. Someone bugging a conference room might overhear that, after a hundred million dollars of research and testing, SwiGLU "worked really well." That single sentence, he said, would be enormously valuable to China or any other rival catching up, saving it research money and time. "It's really, really hard to contain something so small yet so impactful," he said.
SwiGLU is a real and fairly modest example of the kind of finding he meant. Noam Shazeer introduced it in a 2020 paper as a tweak to the feed-forward layers inside a Transformer, the architecture behind today's language models. Instead of passing data through a single standard activation, the layer multiplies two learned projections, one of which first passes through a function called Swish. In Shazeer's tests, with parameter and computing budgets held equal, SwiGLU reached a held-out log-perplexity (a measure of prediction error where lower is better) of 1.636, against 1.677 for the standard ReLU baseline. Gains on follow-up tasks such as question answering were noisy and not universal, and the paper offered no theory for why the change helps. The technique has been public since 2020, so Blundin's $100 million sentence illustrated how small and portable such lessons are; it did not describe anything overheard at OpenAI.
"It's hard to say, Hey, AI, don't let that happen," Blundin said, asking how an AI could defend against something "so slippery."
AIs watching their own conference rooms
Wissner-Gross answered with what he called a "sci-fi thought experiment." If frontier labs end up in a loop of recursive self-improvement, with AIs designing their replacements rather than humans meeting in a conference room, then "the AIs can monitor their own digital conference rooms and make sure that there are no interlopers."
Blundin saw where that led. It would imply that nobody at OpenAI knows what the AI is doing well enough to discuss it in a conference room, which he called "kind of a scary, scary end to that sentence." He imagined the staff's response to a would-be spy: "we don't have any idea what it's doing anyway. So bug us all you want. That's not good."