FBI Arrests OpenAI Engineer as U.S. Agencies Accuse Six Chinese AI Firms of Mass Model Theft

The FBI has arrested a Saudi national employed as an engineer at OpenAI, accused of planting a sophisticated audio recording device inside an executive conference room — a case that landed even as the company had already been quietly training its staff to watch for spies at playgrounds, in bars, and anywhere else strangers might try to pry loose work secrets.

The arrested engineer, identified as Abdulrahman Alfozan, a former Meta employee, is at the center of an active federal case. But his arrest is one thread in a much larger tapestry of espionage targeting the AI industry, one that U.S. intelligence agencies now say has reached industrial scale.

A joint advisory released September 8, 2026, by the National Security Agency, the Cybersecurity and Infrastructure Security Agency, and the FBI identified six China-based AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI — alleging that each had conducted sustained, large-scale efforts to siphon proprietary capabilities from leading American AI models, among them Claude, GPT, Gemini, and Grok.

Even before the Saudi OpenAI employee was arrested yesterday, the startup had been warning staff to beware of people trying to glean work-related information from them — from honeypots to playground spies.— Meghan Bobrowsky, Reporter

According to a Wall Street Journal report, OpenAI's internal security training has included scenarios in which strangers approach employees at San Francisco playgrounds — sometimes using children as props — to extract work-related details through casual conversation. The warnings extend to honey traps and other classic intelligence tradecraft, reflecting how seriously the company has come to treat the threat.

The joint NSA-CISA-FBI advisory describes what the agencies call 'industrial-scale' operations. According to the advisory, the six named Chinese firms drew 'billions of tokens' from millions of interactions with U.S. frontier models; the agencies say the activity stretches back to at least late 2024 and was carried out 'likely with the knowledge of the Chinese government'. As of publication, none of the six companies had publicly addressed the specific allegations, according to ibtimes.co.uk.

The technique at the center of the advisory is known as 'distillation' — querying a more capable AI model at massive scale to generate training data that can be used to build a cheaper, faster rival system. The agencies allege the campaigns were designed to let Chinese AI companies extract capabilities from U.S. frontier models while dramatically reducing the research, computing, and development costs required to build comparable systems independently.

According to the advisory as reported by ibtimes.co.uk, DeepSeek allegedly directed organized distillation campaigns at Claude Sonnet 3.7, Claude Sonnet 4, and Claude Sonnet 4.5, along with several GPT, Gemini, and Grok models, using the output to generate training data for its R1 and V3 models. In a separate set of allegations, Moonshot AI is accused of pulling Claude Fable 5 data to develop its Kimi-K3 system and GPT-4o data to develop Kimi-K2.

To avoid detection, the companies allegedly used fraudulent accounts, bulk purchases of premium subscriptions, and proxy services described as 'transfer stations' — intermediaries that resold access to U.S. models at reduced prices while stripping out identifying metadata. According to reports, the routing infrastructure was built to spread requests across several pathways and shift automatically to alternatives whenever one was blocked.

The September advisory is not the first time these firms have faced such accusations. According to earlier disclosures cited by ibtimes.co.uk, Anthropic reported that three Chinese AI laboratories — DeepSeek, Moonshot, and MiniMax — produced more than 16 million exchanges with Claude by means of roughly 24,000 fraudulent accounts. In an April 2026 memorandum, the White House Office of Science and Technology Policy observed that models produced through unauthorized distillation 'do not replicate the full performance of the original,' even if they can look comparable on certain benchmarks.

The advisory also recommends a countermeasure that raises its own questions: rather than simply blocking suspected bad actors, U.S. AI companies are advised to subtly alter responses — or route suspects to less capable models — without telling those users that anything has changed. For ordinary users, that recommendation surfaces a transparency concern: the same infrastructure that could be used against foreign intelligence operations could, in principle, be applied more broadly.

Taken together — the bugged conference room, the playground warnings, the billions of extracted tokens — the picture that emerges is of an AI industry that has become one of the most actively contested intelligence environments in the world, where the stakes include economic dominance and military applications of frontier technology.

Why it matters — The arrest of an insider and a sweeping government indictment of six Chinese AI firms on the same week illustrates that the race for AI supremacy has become an active espionage war — one whose outcome could shape both economic and military power for years.

Reported by wsj.com, mezha.net, washingtonpost.com, cisa.gov, ibtimes.co.uk, reuters.com