Anthropic red team warns small AI models may soon be able to copy themselves onto cheap smart devices

Anthropic's Frontier Red Team published research on October 12, written by William Zhou and Daniel Freeman. It argues that small AI models may be only months away from being able to copy themselves onto ordinary internet-connected devices, such as smart ovens and off-brand Android TV boxes, without needing a datacenter. In the team's tests, a 9-billion-parameter model got inference running on eight of the nine hardware types tried. Models with 30 billion parameters or fewer can already carry out many of the individual steps needed to replicate, but they make mistakes when they try to chain those steps together. In the team's simplified tests, GLM-5.2 and stronger models could replicate from start to finish. If small models keep improving at their past rate, the authors expect them to reach that level by the end of the year, with small replicators that carry their own weights possible soon afterward. To show that expensive chips are not required, the team linked 15 off-brand Android TV boxes, each costing about $55, and ran the 26-billion-parameter Gemma4-26B-A4B model split across them. The authors say a replicator like this would be hard to shut down because there is no central point to switch off. They also say that, unlike conventional malware, it could work out unfamiliar devices on its own, though it might spread slowly, doubling only every few days.