Anthropic: small AI models could soon copy themselves to TV boxes

Anthropic's Frontier Red Team bought 15 off-brand Android TV boxes at about $55 each, networked them, and ran Gemma4-26B-A4B split across all of them. It worked, at 0.39 tokens per second. A 9B model got inference running on 8 of the 9 hardware types tested. The point of the exercise: an AI no longer needs a data center to run.

Can these models spread on their own? Not yet. Models of 30B parameters or fewer can handle many of the individual steps but make mistakes when chaining them end to end. In a simplified test of 8 emulated Android nodes, GLM-5.2 was the only open-weight model to replicate fully. It reached all 8 nodes in 4 of 5 trials, but the way in was a trivial exploit (an unauthenticated ADB port), so Anthropic calls this an upper bound. If small models keep improving at their historical rate, they should match GLM-5.2 by end of 2026. Authors William Zhou and Daniel Freeman say home and small-business IoT devices would likely be the first targets.

The proposed defense is classic botnet hygiene: ISPs watching for multi-gigabyte model transfers, quarantining infected subscribers, honeypots, and secure defaults from device makers. A replicator spread across devices would have no central host to switch off.

Why it matters: the worm risk is moving from the data center to the cheap box behind your TV.