Anthropic Warns Small AI Models Could Self-Replicate Through IoT Devices

Summary

Anthropic warns that improving small AI models could eventually enable self-replicating malware that spreads through Internet of Things devices. The company tested quantized open-weight models on real or simulated low-cost Android TV boxes and other IoT hardware, distributing model workloads into shards across devices. Gemma4-26B-A4B required 15 boxes with 4GB of memory each to establish inference, while Qwen3.5-9B configured other large language models to run on eight of nine tested hardware types. The models used agents with exploitation tools and followed a four-stage process: finding targets, gaining code execution, establishing inference, and propagating to new hosts. Anthropic says such replicators could adapt payloads to unfamiliar devices and leave prompts, code, and notes for later attempts, potentially making them harder to stop than conventional worms. The experiment did not achieve complete replication: the tested small models could not independently discover devices, exploit them, install a payload, configure inference, and repeat the cycle. Distributed Gemma inference ran at only 0.39 tokens per second, with estimated setup times of 71 to 150 hours for Gemma4-26B-A4B and 29 to 51 hours for Qwen3.5-9B. Anthropic said 30-billion-parameter models can perform many individual steps but still make end-to-end errors, and projected that smaller models could reach similar limited capabilities by the end of the year. The setup relied on Android devices with an unauthenticated Android Debug Bridge exposed, and GLM-5.2 orchestrated the process while a smaller Llama model ran on the targets. Anthropic therefore demonstrated distributed inference and complex agentic setup, not autonomous self-replication. Researcher Xiaohan Fu said the study does not establish that small models would choose to replicate and called for reproducible prompts and instructions.