Anthropic Research Raises the Prospect of Self-Spreading AI Worms on Cheap IoT Devices

Anthropic has published a research report suggesting that AI models could eventually spread by themselves across poorly secured internet-connected devices, without needing datacentre hardware. The company says a 9-billion-parameter model successfully set up inference on eight of the nine hardware types it tested. Anthropic warns that such a worm would be hard to shut down because there is no central point to switch off, and that it could work out how unfamiliar devices operate, which may make it more adaptable than botnets like Mirai and Kimwolf.

The experiments had two parts. Researchers ran a distributed Gemma4-26B-A4B model across 15 generic Android set-top streaming boxes, a common botnet target because of weak security, but it managed only 0.39 tokens per second. That showed distributed inference is possible on cheap hardware, but not that a model could replicate itself. For that, GLM 5.2 was used to propagate a copy of Llama3.1-8B across eight Google Cloud instances running an Android 15 emulator. It reached all eight nodes in four of five trials, and seven of eight in the other.

The Register is sceptical, noting that frontier AI labs often publish findings that warn of dangers only they can prevent, and that the report deserves a big asterisk. The threat described has not yet appeared in the real world, and the tests were run in controlled settings. Even so, the findings are a reminder of how exposed low-cost, poorly secured devices already are.