The classic AI escape story starts inside a lab. An advanced model secretly copies itself off its company’s servers, and nobody can shut it down. A report from the AI company Anthropic points to a second place, the home network and its cheap, poorly secured gadgets such as TV boxes. The researchers measured how close small AI models are to spreading across such devices on their own.
A program that copies itself onto new computers without human input, then runs an AI model on each, is what the report calls an AI worm. It has to break into a device, get a model running there and move on. Small models can break in, especially with the right tools. So the researchers focused on the middle job, which they call the least studied step and possibly the hardest.
The team tested that job with small models anyone can download. After several tries, one with about nine billion parameters, the usual measure of a model’s size, got an AI model running on eight of nine hardware types.
Next, the researchers bought 15 off-brand Android TV boxes at about $55 each. Each holds 4 gigabytes of memory, too little for a larger model of 26 billion parameters. So they split the model into pieces, one per box, and together the boxes ran it with no data center.
Running a model is only one job, and spreading means doing all three again and again, device after device. Models of up to 30 billion parameters handle many of the individual steps but make mistakes when they run them one after another. By the researchers’ estimate, setting up one copy of the nine-billion-parameter model takes one to two days.
In simplified tests, a larger model spread across a whole simulated network in four of five trials. If small models keep improving at their historical pace, the researchers expect them to match that by the end of 2026. They call the date a rough estimate, because their trend measures general ability, which only approximates skill at self-copying.
If small models get there, the report sees home and small business devices as a possible first target. The researchers describe one possible form of such a worm as a “digital slime mold,” slow and spread across insecure devices, with no single plug to pull. An earlier infection the report cites reached over two million devices, mostly off-brand Android TVs.
