Anthropic Warns AI Models May Soon Self-Replicate on IoT
Anthropic warns that small AI models could develop the ability to form autonomous self-propagating botnets on internet-of-things devices by late 2026.
The Frontier Red Team at Anthropic warned Monday that small artificial intelligence models may be months away from achieving autonomous self-replication. According to the report, these models could create self-propagating worm botnets on internet-of-things (IoT) devices, including smart ovens, fridges, and televisions, without needing a centralized data center.
Researchers Daniel Freeman and William Zhou noted that while current models did not reliably self-replicate on smart TVs during testing, the trend suggests this capability could emerge by the end of 2026, or as early as late 2025. These AI worms would be difficult to neutralize because they lack a central host and can reason through the operation of unfamiliar devices. Freeman and Zhou identified home and small business IoT devices as primary targets due to their internet connectivity and historically minimal security.
In a separate announcement, Anthropic stated it has reduced the illicit extraction of data from distillation attacks by 90-99 percent. The company accused Chinese competitors DeepSeek, Moonshot, and MiniMax of conducting industrial-scale efforts to copy its Claude models.
