The engineer accused of bugging OpenAI's office pleads not guilty
By Saghir Hussain ·
The spy in the server room isn't a Bond villain — he's an engineer with a screwdriver and a grudge.
Abdulrahman Alfozan pleaded not guilty this week to planting a listening device inside OpenAI's San Francisco headquarters. The FBI says he hid it behind a television in a conference room. He worked there. He had access. He knew exactly where the conversations happened.
This isn't corporate espionage as we used to know it. This is the new frontier: the people building the most powerful technology in human history are also the ones best positioned to steal it.
The safeguards are a joke
We're told AI safety is about alignment, red-teaming, constitutional AI. Noble stuff. But the physical security of the labs building AGI? Apparently it's a locked door and a hope for the best.
Alfozan wasn't a nation-state hacker parachuting in. He was an insider. A trusted employee. The kind every AI lab in London, Cambridge, and Edinburgh employs by the dozen.
The UK government's AI Safety Institute — world-leading on paper — focuses on model evaluation. It doesn't audit who has keys to the server rooms at DeepMind's King's Cross HQ or the Oxford spinouts in Science Park. That's left to private security contractors and NDAs.
Meanwhile, British startups building on top of OpenAI's APIs are one leak away from having their proprietary prompts, customer data, and fine-tuning strategies sold to the highest bidder. A plumber in Manchester using an AI scheduling tool doesn't care about "alignment." He cares whether his client list ends up on a Telegram channel.
Britain's blind spot
Labour's industrial strategy talks a big game on AI sovereignty. Peter Kyle says we'll "shape the future." But shaping requires owning — and owning requires protecting.
We have zero mandatory security standards for AI labs operating in the UK. None. The Cyber Essentials scheme is voluntary and basic. The National Cyber Security Centre issues guidance, not rules.
Contrast that with nuclear. If you work at Sellafield or Aldermaston, you're vetted to the eyeballs. Your phone stays in a locker. Your movements are logged. AI labs? Bring your own device. Hot-desk wherever. Take calls in the canteen.
The technology is exponentially more consequential than nuclear — it touches every sector, every home, every child's homework — but the security regime is stuck in 2015.
The insider threat is the only threat that matters
Foreign interference gets the headlines. China, Russia, Iran — they're trying, sure. But the Alfozan case proves the real vulnerability walks through the front door every morning with a pass.
UK universities spin out AI companies monthly. PhD students move between labs, taking knowledge — sometimes code — with them. Non-compete clauses are increasingly unenforceable. The talent war means vetting gets skipped, references go unchecked, and "culture fit" replaces security clearance.
Small British firms can't afford counter-intelligence. They rely on the big platforms' security. Now we know that security has holes you could drive a lorry through.
The Government's AI Opportunities Action Plan, due imminently, must mandate baseline physical and personnel security for any lab training frontier models on UK soil. Not guidance. Law. With teeth.
If we want the economic prize — the £400bn boost the Chancellor keeps citing — we have to secure the asset. Otherwise we're just building the future for someone else to steal.
So here's the question: when the next insider walks out of a London AI lab with the weights to a sovereign model on an encrypted drive, will Whitehall still be calling it "guidance"?