Fifteen cheap TV boxes ran one AI model between them

Fifteen off-brand Android TV boxes, one AI model split across all of them, and it ran. That was the experiment Anthropic's Frontier Red Team published on 13 October, and the question behind it was simple: could a language model survive outside a data centre, on the cheap and badly maintained hardware already sitting in people's living rooms?

Close, is the answer. They got Gemma4-26B-A4B distributed across those fifteen boxes. A smaller model managed to install the software it needed on eight different types of hardware 84 percent of the time. What they did not build is a worm. Nothing self-replicating is loose, the test devices had debugging access left wide open in a way real ones usually do not, and at current speeds one device would need somewhere between 28 and 150 hours to do its share of the work. The researchers call it signs of life, and reckon models of 30 billion parameters or fewer could be capable enough by the end of this year.

The reason this lands at home has little to do with AI. The devices in question are the ones nobody updates: the streaming box bought from a marketplace seller, the camera above the door, the router from two providers ago. Mirai has been recruiting exactly that population for a decade, and BADBOX 2.0 arrived on Android boxes already compromised, straight out of the factory.

Which means the advice is the old advice, and that is genuinely reassuring. Buy from makers who still ship firmware updates, keep smart devices on a guest network away from your PC and phone, and unplug anything you cannot account for. Tendvane's network scan lists what is actually connected to your home network, which tends to be a longer list than people expect.