Anthropic warns AI may self-replicate across IoT devices

Anthropic’s Frontier Red Team warned that small AI models could become capable of autonomously copying themselves across poorly secured internet-connected devices within months or by the end of 2026. Tests found that models with 30 billion parameters or fewer could perform many steps in the process, but current systems struggled to complete it; only GLM-5.2 and stronger models succeeded in a simplified test with security barriers removed. Researchers also distributed an open model across 15 low-cost Android TV boxes to demonstrate that inexpensive hardware can run AI, though it ran very slowly. Anthropic said a self-spreading AI worm could be difficult to shut down because it would lack a central host, but stressed that replication would likely be slow and that no reliable, fully autonomous replicator has been demonstrated. The company urged stronger security for IoT devices, which have historically been vulnerable to compromise.