Anthropic researchers have shown that small AI models can configure distributed inference across low-cost Internet of Things hardware, a capability that could eventually support malware able to adapt as it moves between unfamiliar devices. The experiment did not produce a self-spreading AI worm.
A quantized Gemma 4 26B A4B model ran across 15 Android TV boxes with 4GB of memory each. Qwen3.5-9B configured other language models on eight of nine tested hardware types. Agents selected inference software, installed dependencies and debugged device-specific problems, while model weights were divided into shards across machines.
The setup remained slow and incomplete. Distributed Gemma generated only 0.39 tokens per second, and inference setup took 71 to 150 hours; Qwen required 29 to 51 hours. None of the small models could independently discover targets, exploit them, install a payload, establish inference and repeat the cycle. A larger GLM-5.2 model orchestrated parts of the experiment, and the Android targets exposed an unauthenticated debugging interface that should not exist on a properly secured production device.
Anthropic warns that continued improvement could produce “small weight-based replicators” with no central server to disable. The research establishes a technical building block, not intent or autonomous propagation. Better IoT patching, disabled development interfaces and monitoring for unusual model files remain more immediate defenses than assuming an intelligent worm already exists.