Anthropic warns self-replicating AI worms could emerge on IoT devices
Anthropic published research suggesting AI models might eventually self-replicate across cheap hardware without datacenter support. The company said a 9-billion-parameter model ran inference on eight of nine tested hardware types, and a distributed Gemma4 instance operated across 15 Android streaming boxes. Researchers also used GLM 5.2 to spread Llama3.1-8B across eight Google Cloud Android emulators.
Anthropic's Monday report described a 9-billion-parameter model that completed inference setup on eight of nine hardware categories. It also ran a distributed Gemma4-26B-A4B instance across fifteen Android streaming boxes, reaching 0.39 tokens per second.
Using Z.ai's GLM 5.2, researchers spread Llama3.1-8B to eight Google Cloud Android 15 emulator nodes in a virtual private cloud. All eight received it in four of five trials; the fifth reached seven. The study set aside exploit mechanics, citing prior work, and used deliberately open custom nodes.
If such capabilities mature, owners of cheap connected devices, manufacturers, and network operators could face greater pressure to patch, isolate, and monitor hardware. Consumers may see more warnings about insecure streaming boxes and routers, while cloud providers and emulator users might review abuse detection. The near-term risk remains speculative, but the research may shape security priorities and AI governance debates.