OpenAI is training staff to spot spies, from playground strangers to honey pots

OpenAI has been warning employees about attempts to steal company secrets through offers of money, eavesdropping in cafes, or romantic approaches. Staff at Anthropic and Google have received similar warnings about competitors and state-sponsored actors, The Wall Street Journal reported.

Those warnings predated the arrest of OpenAI engineer Abdulrahman Alfozan, whom prosecutors accuse of hiding a recording device in an executive conference room. Alfozan, a 35-year-old Saudi national, was arrested on Oct. 9 and appeared in federal court on Oct. 13. Prosecutors have not alleged that he acted for a foreign government or identified an intended recipient for the recordings.

What OpenAI tells its staff

In one training scenario, a spy who regularly brought a child to a San Francisco playground tried to befriend an OpenAI employee there, using small talk to ask about the employee's work.

OpenAI tells employees not to wear anything with the company's logo outside the office. Staff are told to use screen protectors that limit who can see their screens on public transit, and to wipe down whiteboards when they finish with them, in case kitchen or cleaning staff have been compromised. Some sessions also cover "honey pots," agents who use sex or romance to get close to a target.

The company has a security team with experience in physical security, insider risk, and information security, an OpenAI spokesperson said.

What the complaint alleges

OpenAI staff found the first device on or about Sept. 30, plugged into a USB port on the back of a television in an executive conference room in San Francisco. It was wrapped in toilet paper, and its contents were encrypted. OpenAI determined it had been installed on or about Sept. 28, the criminal complaint alleges.

Alfozan had no authority to enter that room or attend its meetings, but security cameras showed him near it on several days before the device was found, the complaint says. Before Sept. 28, he used a productivity tool to ask about OpenAI's meeting rooms, room calendars, and "leadership and exec" locations, prosecutors say.

Between Sept. 24 and Sept. 30, a chatbot account tied to Alfozan's spouse submitted prompts about "build[ing] a continuous audio recorder" and "encrypt[ing] the SD card." Prosecutors say that account shared multiple identifiers with Alfozan's own chatbot account. On Oct. 5, Alfozan went into the conference room and appeared to look behind the TV. Also on or about Oct. 5, OpenAI security found a second, similar recording device in a bag he had left in another room.

The charge is using or trying to use a device to intercept oral communications, which carries up to five years in prison and a $250,000 fine. His attorney called the allegations "wholly out of character" and noted that Alfozan is not charged with stealing trade secrets or acting as a foreign agent.

OpenAI said it fired Alfozan. "At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data," the company said.

Why AI labs make attractive targets

Employees change jobs often across the AI industry. Apple is suing OpenAI over claims that former Apple employees took confidential material with them, and Elon Musk's xAI made a similar claim in a lawsuit against an engineer who joined OpenAI. In January, a federal jury convicted former Google engineer Linwei Ding of economic espionage and theft of trade secrets tied to Google's AI technology.

"AI companies are scaling their workforces at unprecedented speed. The question is whether their ability to manage insider risk is scaling at the same pace," said Aviv Nahum, chief executive of Above Security, a startup that uses AI to detect insider risks.

Phil Venables, a former Google Cloud chief information security officer, is a partner at Ballistic Ventures, which invests in Above Security. He said the concentration of so much value in a few AI companies makes them bigger targets. "The resources and money required to train foundation models makes them even more valuable intellectual property," he said.