Anthropic Researchers: Small AI Models May Soon Self-Replicate on IoT Devices

Huo Xing Caijing reports that researchers William Zhou and Daniel Freeman from Anthropic’s Frontier Red Team published a study on October 12, 2026. The paper states that small models may acquire the ability to self-replicate on commercial IoT devices such as smart ovens and generic Android TV boxes within months. These replicas lack a single kill switch, may double in number on a daily timescale, and could autonomously adapt to unfamiliar devices. Evaluations show that a 9B model successfully performed inference on 8 out of 9 hardware types. Models up to 30B have already accomplished many individual steps of replication but fail when combining them end-to-end; in simplified evaluations, GLM-5.2 and stronger models completed full end-to-end replication. The team also distributedly ran Gemma4-26B-A4B across 15 generic Android TV boxes costing approximately $55 each, achieving a speed of about 0.39 tokens/second per device, with each box holding two layers of the model. End-to-end testing was conducted across eight GCP nodes behind a firewall, propagating via unauthenticated ADB ports and deploying Llama 3.1-8B inference. Among all open-source models tested, only GLM-5.2 completed full replication, successfully spreading to all eight nodes in four out of five trials. The paper defines autonomous replication and adaptation as a four-step process: identifying targets, gaining execution, establishing inference, and continuing propagation from new hosts, and concludes that no fully autonomous, runaway replicator currently exists.