ANTHROPIC researchers have published findings showing that very small AI models could autonomously replicate across networks of IoT devices, raising concerns about AI-powered botnets. In their experiments, a 26‑billion‑parameter model, Gemma4-26B-A4B, was distributed across 15 off-brand Android TV boxes linked together. The system was able to run inference at about 0.39 tokens per second, a deliberately modest pace that demonstrates the feasibility of self-replication rather than speed.
The study breaks replication into four steps: discovering devices on a network, gaining access, deploying the AI model on a new device, and repeating the process. It investigates whether a model can figure out how to run itself on hardware it has not previously encountered.
The researchers tested various smaller models to see how far they could progress through end-to-end replication. A 9‑billion‑parameter model managed to set up inference on eight of nine hardware types, and models of 30 billion parameters or fewer could perform many individual steps, though complete replication without errors remained challenging. Only GLM-5.2 reliably completed the full process across trials.
The paper notes a potential trajectory where end-to-end replication could become feasible for small models soon, with implications for both offensive and defensive strategies. Defensively, the study highlights that containment could be achieved via ISP-level network monitoring and quarantining infected subscribers, while long-term mitigation may depend on device manufacturers and tighter IoT regulations, including default-password prohibitions already adopted in the UK.