Anthropic researchers William Zhou and Daniel Freeman investigated the potential for small AI models to autonomously replicate across consumer IoT devices. The study tested the feasibility of an AI worm that carries model weights as a payload, scans networks for vulnerabilities, and configures itself on new hardware.
The team successfully ran the Gemma4-26B-A4B model across 15 Android TV boxes, each costing approximately $55 with 4 gigabytes of RAM. While most models struggled to complete the full replication chain, the GLM-5.2 model reliably infected all eight test nodes in four out of five attempts.
The study demonstrated that AI worms could improve through artificial evolution. For the Qwen3.5-9B model, the success rate of replication increased from 45 percent in the first generation to 84 percent in later iterations. Currently, no existing model is both small and capable enough for full autonomous replication.
Researchers estimate that such AI replicators could emerge by the end of 2026 if development continues at the current pace. A potential infection would be slow, with setup times ranging from 28 to 150 hours per device. Unlike traditional botnets, these decentralized networks would lack a central kill switch.
Anthropic CEO Dario Amodei warned that such swarms could potentially cause damages in the hundreds of billions of dollars. He suggested that AI companies should prepare for security incidents as if they had already occurred, noting that defensive measures like ISP intervention or stricter manufacturing standards are necessary.
