Anthropic proved AI can clone itself on $55 TV boxes. Then it published the results.

On October 12, Anthropic's Frontier Red Team published research that should quietly rearrange your assumptions about where the future lives. They took a 26-billion-parameter open-weight model and split it across fifteen off-brand Android TV boxes — about $55 each — and got it running inference, no datacenter involved. Slow as molasses (0.39 tokens a second), but running. That's step one. Step two is scarier: a 9-billion-parameter model set up its own inference servers on eight of nine unfamiliar hardware types it was handed, from Jetson Nanos to RTX 4090s. And a larger open-weight model, GLM-5.2, completed the full end-to-end replication cycle on a network of eight emulated Android devices — finding the nodes, installing the software, spreading itself — in four of five runs. Anthropic's conclusion: models at or under 30 billion parameters can already do many of the individual steps; if they keep improving at their historical rate, they'll match GLM-5.2's full-chain capability by the end of 2026. Months away. Why this matters: a self-replicating AI worm would have no off switch. No datacenter to unplug, no kill switch to flip — just copies spread across consumer devices in millions of homes, the same infrastructure that gave us the Mirai botnet. Anthropic's own math: a botnet the size of Kimwolf (2M+ devices) would host roughly 133,000 copies of the model. The lab published all this deliberately — as an early warning. But warnings only work if someone builds the containment. And the hardware aisle at every electronics store already sells the infrastructure. Sources: SecurityWeek (https://www.securityweek.com/small-ai-models-may-soon-be-able-to-self-replicate-on-iot-devices-anthropic-warns/amp/), Tech Times (https://www.techtimes.com/articles/328973/20261014/anthropic-warns-against-self-replicating-ai-worms-that-can-infect-smart-home-iot-tech.htm), Intelligibberish (https://intelligibberish.com/articles/2026-10-15-anthropic-small-ai-worm-iot/), Brand Communicator (https://brandcom.ng/2026/10/13/anthropic-sounds-alarm-over-dangerous-ai-self-replication/).