Anthropic Warns Small AI Models Could Self-Replicate on IoT Devices Within Months

Anthropic released experimental findings on Monday, warning that AI models with fewer than 30 billion parameters could gain the ability to self-replicate on commercial IoT devices within months, including smart ovens and Android TV boxes.

In the experiment, Anthropic tested whether AI models could discover vulnerable devices, copy themselves onto those devices, and keep spreading without human intervention. Models under 30 billion parameters could complete several steps of the replication process but struggled to finish the entire chain on their own. In a simplified test environment with some safety checks removed, stronger models such as GLM-5.2 achieved end-to-end self-replication.

Anthropic cautioned that such self-replicating models would be difficult to shut down, with no single kill switch. Unlike traditional non-AI malware, they can understand unfamiliar devices on their own and may even target data center processors. The research focused on Android TVs and smart home appliances because these connected devices generally lack strong security protections.

The company also said that if small models continue to improve at the same pace, autonomous replication could be possible before the end of 2026, followed by small replicas derived from model weights. These autonomous AI systems would likely spread slowly and rely on widely available, weakly protected computing resources. The report recommends stronger system defenses to make it harder for such AI agents to propagate.