On October 9, the FBI arrested an OpenAI engineer and charged him with hiding a recording device in one of the company's executive conference rooms.

Prosecutors say Abdulrahman Alfozan, a Saudi national who joined OpenAI's product team in July 2025, plugged the device into a USB port on the back of a television in OpenAI's San Francisco office. OpenAI found it, had its own models analyze it, and says it had "advanced encryption, audio monitoring capabilities, and a mechanism for transferring data designed to evade monitoring." A camera OpenAI put in place allegedly caught him looking behind the TV. He was fired and arrested about a week after the discovery, and faces up to five years in prison if convicted.

FBI Director Kash Patel said the aim was to "capture sensitive communications about the company's super intelligence technology." OpenAI says it has "no indications" that he accessed or took model weights, source code or customer data. Nobody has said why he did it, officials have not alleged that any government was involved, and he has not been convicted of anything.

Wispr Flow Notetaker brings your meetings into Claude or ChatGPT through a built-in connector. Ask for a proposal draft or a deal recap, and your AI works from an accurate record of what was said.

Get Wispr Flow Notetaker

It's the most cinematic case so far. It isn't the only one. Over the past year, the fight over AI secrets has spread in four directions: out the door with departing employees, through the API, into the inboxes and hiring pipelines that state actors work, and through the chip supply chain.

xAI says Xuechen Li copied confidential Grok material from his work laptop on July 25, 2025, the same day he finished selling xAI stock worth about $7 million, then resigned three days later to join OpenAI. xAI says he admitted it in a handwritten document. Weeks later, xAI sued OpenAI itself for poaching staff to get at its secrets. A federal judge dismissed that suit in February, writing: "Notably absent are allegations about the conduct of OpenAI itself." xAI says it will appeal, and OpenAI is trying to recover more than $1 million in legal fees.

Scale AI sued Mercor in September 2025, alleging that a sales employee took more than 100 confidential documents and pitched one of Scale's biggest customers before he had left. Mercor's co-founder said the company never used Scale's data and offered to have the files destroyed.

The biggest case is Apple's. In July, Apple sued OpenAI, its hardware chief Tang Tan and former Apple engineer Chang Liu. Apple says OpenAI used job interviews to pull out hardware secrets, asking candidates to bring components and "design artifacts" to show-and-tell sessions, and that Liu kept access after he left and downloaded confidential files. OpenAI denies all of it. On October 14, a federal judge in San Jose heard Apple's request for an injunction and OpenAI's motion to dismiss.

Read the defenses and a pattern shows up. The accused don't only deny the theft. They argue the lawsuit is really about hiring. OpenAI says Apple is trying to stop its people from leaving and is treating ordinary job interviews as theft. For anyone recruiting senior talent, that is the stake: a trade-secret suit is now one of the ways a company tries to slow down the people walking out its door.

You can just ask it questions. Distillation means training a cheaper model on a stronger model's answers. In February, OpenAI told the House China committee that DeepSeek was making "ongoing efforts to free-ride on the capabilities developed by OpenAI and other U.S. frontier labs," using obfuscated third-party routers to hide where the requests came from. Eleven days later, Anthropic published its own numbers.

None of this has reached a courtroom. The labs accused are in China, Anthropic doesn't sell there, and the traffic came through proxy networks. So the response is technical and political instead: classifiers that spot extraction, tighter account checks, indicators shared with other labs, and briefings for Congress.

At a House Intelligence Committee hearing in July, Frank Cilluffo of Auburn University's McCrary Institute said that if you're China's Ministry of State Security today, "you're looking at frontier model companies." H.R. McMaster said the counterintelligence response "has lagged way behind."

The cases back that up. In January, a jury convicted former Google engineer Linwei Ding of stealing AI trade secrets, after he uploaded more than 500 confidential files on Google's AI supercomputing infrastructure to a personal account while lining up roles at two Chinese AI startups, one of which he founded. A judge later threw out the economic espionage conviction, and in September he was sentenced to just under a year in prison.

Most approaches are quieter than a device behind a TV. In October, Proofpoint reported that a China-linked group it tracks as TA419 had posed as former US officials to phish AI policy experts, and earlier this year had impersonated an Anthropic employee to target a think-tank researcher. The same week as the OpenAI arrest, the Justice Department said North Korean remote IT workers had been hired by more than 70 US companies. The department didn't name the companies, but the method works on any company hiring remote engineers in a hurry. And the labs' own threat reports now read like intelligence briefings: on October 8, OpenAI described shutting down a Russian operation run through a fake think tank and an Iranian one built on invented journalists.

The other way to catch up is to get the chips. In March, prosecutors charged Super Micro co-founder Wally Liaw and two others with routing Nvidia-powered servers to China through shell companies in Southeast Asia. Super Micro, which was not charged, said it was cooperating, and its shares fell by about a third. Earlier this month, Greg Lui, the owner of a California server company, was charged with smuggling more than $300 million of export-controlled servers with Nvidia chips to China.

The prize got bigger. OpenAI's August share buyback valued it at $852 billion, and Anthropic is reportedly preparing to go public. A model's weights, training recipes and hardware plans are now some of the most valuable files in the world.

People move constantly, and they leave rich. In Issues 02 and 04 we tracked co-founders leaving Thinking Machines and humans&. Every move is a chance for files to move too, and the xAI case allegedly happened in the same week as a share sale and a resignation.

Copying got cheap. Distillation needs accounts, not agents. Anthropic's numbers suggest it can be run at the scale of millions of requests.

The labs spent two years bidding for each other's people. Now they're checking what those people carried in, and what they carried out.

The scale of the problem, from the cases in this issue.

Cases described here are allegations unless a court has ruled, and everyone charged is presumed innocent. Other sources: Nextgov on the July House Intelligence Committee hearing; Proofpoint via Security Boulevard; OpenAI threat report, October 8, 2026; CNBC on Super Micro.

Nothing here is legal or investment advice.