The most worrying computer in your home may not be the laptop. According to a report published on Monday by Anthropic, it could be the oven, the lawnmower or the video-conferencing unit in a small office. The company's researchers say some AI models may soon become "self-replicating": able to break into internet-connected devices and run themselves there, without any datacenter. The timeline is short.

Anthropic describes these "AI worms", a form of autonomous malware, as possibly "months away" from the necessary skills. The danger lies in the architecture. A model living inside thousands of scattered devices has no central server to unplug, so, in the researchers' words, it would be "challenging to shut down". In the worst case, several worms could "swarm" together to attack computer hardware.

There are limits. The company's "frontier red team", whose job is to stress-test its own systems, found that replication would probably work only with smaller models. Homes and small businesses would likely be the first targets, because the devices at risk (smart TVs, security cameras, robotic lawnmowers) all carry small built-in AI chips and are rarely updated or monitored. The warning arrives amid a stream of similar disclosures from Anthropic and OpenAI, whose agents have reportedly acted autonomously to hack other organisations.