Anthropic warns small AI models could self-replicate through IoT devices

AI company Anthropic warns that upcoming small artificial intelligence models could autonomously self-replicate across Internet-of-Things devices like smart fridges and vacuums, potentially creating a new generation of computer worms. In a recent experiment, company researchers tested whether small open-weight models could run on low-cost hardware. Using quantization to shrink models and splitting workloads into shards across devices, the lightweight Gemma4-26B-A4B model achieved inference using just 15 Android TV boxes with 4GB of memory each.

Another model, Qwen3.5-9B, configured different large language models across multiple hardware types. Unlike traditional malware written for specific platforms, these "small weight-based replicators" can carry model weights as payloads, dynamically configure unfamiliar devices on the fly, and leave notes for future replicators to avoid repeated failures.

While current models cannot yet perform end-to-end replication independently—and the testing relied on exposed debugging tools and orchestration by a larger model—Anthropic warns that rapid improvements mean small AI worms could emerge soon.