Anthropic warns small AI models could soon spread themselves across TV boxes
Researchers ran an open model on 15 cheap TV boxes. Small models could be able to copy themselves without a data centre before the year is out – albeit slowly.
Symbolic image · Photo: TGspot.co.il / Wikimedia Commons, CC BY-SA 4.0 (cropped and resized)
Anthropic's Frontier Red Team published a study on 12 October suggesting that small, freely available AI models may be only months away from being able to spread on their own across cheap connected devices – without any data centre. Authors William Zhou and Daniel Freeman ran the open model Gemma4-26B-A4B distributed across 15 Android TV boxes costing about 55 dollars each, though at only around 0.3 to 0.39 tokens per second.
In a contained test network, the open model GLM-5.2 copied itself to all eight nodes in four out of five trials. Models with up to 30 billion parameters can handle many individual steps but often fail to chain them together. "If small models improve at their historical rate, they'll match this capability level by the end of this year," the researchers write. Such a swarm would be hard to shut down because there is no central off switch, but it would spread slowly.
Anthropic stresses the limits: the hacking step in the test was deliberately kept trivial, so the results are an upper bound. As a countermeasure the authors point to quarantine by internet providers: with the earlier Mirai malware, this cleared 92 percent of infections within 14 days. They expect quarantine to be similarly effective as an initial defence against an AI replicator.