Back in April I wrote that CISOs should start treating frontier AI models as an AI-native insider threat. The last few weeks were a good reminder that the old-fashioned kind hasn’t gone anywhere… one allegedly showed up with a microphone wrapped in toilet paper.
Concern #1: The recon was sitting in his own tools
According to the DOJ, OpenAI engineer Abdulrahman Alfozan was arrested Oct 9 and charged with one count of using, or trying to use, a device to intercept oral communications. Per the complaint, around Sept 30 OpenAI found a recording device plugged into a USB port on the back of a TV in an executive conference room. It had allegedly been installed around Sept 28, wrapped in toilet paper, with its contents encrypted. WIRED reports it was a dev board that retails for about $14.
These are allegations, and he has not been charged with espionage. The government hasn’t said whether he acted alone (per the Washington Post, via Slashdot). OpenAI said it has no indications he accessed or exfiltrated model weights, source code or customer data (CNBC).
What jumped out at me was the prep work. Before the install, he allegedly used an internal productivity tool to ask about meeting rooms, room calendars, “leadership and exec” locations and partner meetings. A chatbot account linked to his spouse allegedly asked how to build a continuous audio recorder, improve live listening and encrypt the SD card. That’s a legitimate user doing lookups that make zero sense for his role. Would anyone in your org notice an engineer pulling exec calendars?
Concern #2: A badge isn’t a reason to trust the room
Per the DOJ, he had no authority to access that room. CCTV allegedly showed him near it on multiple dates, and on Oct 5 he allegedly went in and looked behind the TV. That day security also found his unattended bag with his wallet, his ID and a second, similar device.
It was found within about two days, which is decent detection. Nothing stopped it from going in, though. We spend a lot of time in Zero Trust conversations saying that being on the network isn’t a reason to trust you. Same goes for the room you badged into… and that TV’s USB port.
Concern #3: Most insiders aren’t spies
The other story is less cinematic and more relevant to most of us. Nottingham University Hospitals NHS Trust has now dismissed 15 staff, up from 11 in May, for inappropriately accessing the medical records of the three people killed in the 2023 Nottingham attacks (BBC). Doctors and nurses were among them.
The trust is only now introducing a system to monitor record access and flag suspicious activity. It’s in testing, with go-live expected early next year (National Health Executive). The medical director put it plainly: “There is simply no excuse for accessing the records of a patient without a legitimate reason” (Sky News).
Perhaps a better framing: stop thinking of insider threat as “catching the spy.” The spy is rare. The curious nurse, the teller looking up a neighbor’s balance, the engineer pulling exec calendars… that’s the everyday stuff, and it’s sitting in logs you already have. If nobody reviews them, the honor system is your control.
Where I’d start:
- Pick your most sensitive records (customer accounts, patient charts, exec calendars) and actually review who’s looking at them. Alerts on lookups that don’t fit someone’s job beat another policy acknowledgment.
- Treat exec and board rooms like a sensitive system. Real access control, and occasionally check what’s plugged into the AV gear.
- Run an insider scenario at your next tabletop. “The badge worked and the login was valid” makes for an uncomfortable inject, which is the point.
- Tell people you monitor.
Authorized and appropriate aren’t the same thing. Your logs know the difference, but only if someone reads them.
