Anthropic Tests Highlight Potential for Self-Replicating AI Worms on IoT Devices

Anthropic’s Frontier Red Team demonstrated capabilities that could eventually enable small AI models to spread across compromised smart TVs and other IoT devices without centralized inference infrastructure. In controlled tests, a 9-billion-parameter model established inference on eight of nine hardware types, and researchers demonstrated slow distributed inference across inexpensive Android TV boxes. Separately, GLM-5.2 completed a deliberately simplified replication cycle across eight emulated Android devices in four of five runs using unauthenticated Android Debug Bridge access. That experiment relied on an external inference server and researcher-initiated copies; no complete rogue replicator or active AI malware outbreak was demonstrated.

Anthropic estimates that, if historical capability improvements continue, models with 30 billion parameters or fewer could reach GLM-5.2’s tested capabilities by the end of 2026, potentially enabling small self-replicating models afterward. The company characterizes its controlled results as an upper bound on real-world success. Decentralized AI worms could lack a central shutdown point, but limited compute, slow replication and conspicuous model transfers would constrain them and create detection opportunities. Organizations should disable exposed Android Debug Bridge access, eliminate default passwords, isolate consumer IoT devices from business networks, and monitor large model downloads and unusual device-to-device traffic.