Anthropic warns small AI models may self-replicate on IoT devices

In a blog post, Anthropic warned that as small AI models grow more capable, a new kind of self-replicating malware could emerge, infecting smart fridges, robot vacuums and other IoT devices. The company ran small open-weight models on real or simulated low-cost, no-brand Android TV boxes, achieving inference by quantizing and compressing the models and splitting workloads into shards across devices. The lightweight Gemma4-26B-A4B ran on just 15 boxes with 4GB of RAM each, while Qwen3.5-9B could configure other LLMs to run on eight of nine hardware types. Unlike traditional malware, these models can use their own weights to dynamically generate configurations tailored to different devices. Anthropic describes this agent as capable of learning on the job: each failed attempt to compromise a device type leaves prompts, code and notes for later copies to reference. However, no lightweight model could complete end-to-end replication, and speeds were extremely slow — Gemma4 spread across 15 TV boxes managed only 0.39 tokens per second, with a full inference environment taking 71 to 150 hours to set up. For teams working on IoT firmware and edge devices, the experiment exposes two direct risks: many cheap devices ship with an unauthenticated Android Debug Bridge, effectively opening full control of the device; and the barrier to distributed inference has dropped low enough to be met with consumer-grade hardware. The researchers also noted that the replication process was in fact orchestrated by GLM-5.2, so it does not prove that small models would autonomously choose to replicate.