The FBI has arrested an OpenAI engineer accused of hiding a recording device in the company’s San Francisco office. In separate news, Anthropic warned that small AI models could soon copy themselves across poorly secured smart devices.
The two stories are not connected. Both deal with security issues facing the artificial intelligence industry.
FBI Charges OpenAI Engineer in Recording Device Case
Federal agents arrested Abdulrahman Alfozan, a 35-year-old Saudi national, on October 9. The Justice Department announced the case on October 13.
He is charged with using or attempting to use an electronic device to intercept oral communications.
According to the criminal complaint, OpenAI security staff found a hidden audio device on September 30. It was plugged into a USB port on the back of a television in an executive conference room and covered with a small piece of toilet paper.
Investigators believe it was installed around September 28. The room was used for meetings about research, products, business plans, and government work.
The device reportedly had encryption and audio monitoring features. It also had a way to send data that was designed to avoid detection.
Prosecutors say a chatbot account linked to Alfozan’s spouse was used between September 24 and 30 to ask how to build, install, test, and encrypt a recording device.
Authorities also allege Alfozan returned to the room on October 5 and appeared to check behind the TV. A second similar device was found in a bag left in another room.
OpenAI cut his access and later fired him. The company said it has “no indications” he accessed or took model weights, source code, or customer data.
Anthropic Warns of Self-Copying AI on Smart Devices
In a research report published October 12, Anthropic said small AI models “may be months away” from being able to copy themselves onto internet-connected devices such as smart ovens and cheap Android TV boxes.
The company said such a system could be hard to shut down because it would have no single control point. Anthropic described this as an emerging risk, not an active threat.
In one test, researchers ran Google’s Gemma4-26B-A4B model across 15 low-cost Android TV boxes. This showed weak hardware could work together to run a model, but it was not a spreading worm.
In another test, the GLM-5.2 model spread across eight emulated Android devices on Google Cloud in four of five trials. The setup used an unprotected debug port, which made it much easier than a real-world attack.
Anthropic estimated that models with 30 billion parameters or fewer could reach this level by the end of 2026 if progress continues. It noted the process would be slow, which may give defenders time to respond.
The company suggested watching for large AI model file transfers and unusual traffic between devices. It also advised device owners to update firmware and change default passwords.
In the OpenAI case, Alfozan remains in federal custody. He faces up to five years in prison and a $250,000 fine if convicted, and his detention hearing is set for October 19.
