Anthropic research suggests small AI models may self-replicate on IoT devices

The study evaluated various open-weight models on their ability to set up inference servers on unfamiliar hardware. While smaller models like Qwen3.5-9B could manage individual steps, only GLM-5.2 successfully completed the full end-to-end replication cycle in a controlled environment. The researchers estimate that if small models continue improving at their historical rate, they will match this capability level by the end of 2026. Unlike traditional malware, these AI replicators could adapt to new devices and potentially use distributed compute resources to sustain operations without centralized control.

The report highlights that IoT devices, such as smart TVs, are attractive targets due to their low cost and lack of security features. By sharding model weights across multiple devices, a worm could achieve significant computational power. The authors note that while the spread would be slow, with a doubling time of days, the persistence and adaptability of AI-driven malware present a new challenge for cybersecurity defenses.