Anthropic's Frontier Red Team said small AI models could soon self-replicate across internet-connected devices such as smart TVs, according to a report published on the company's website. The company said this could happen within months and without needing a data center.
Researchers bought 15 off-brand Android TV boxes with 1 to 4 gigabytes of RAM for about $55 each and ran Gemma4-26B-A4B, an open-weight model, split across them using pipeline parallelism, reaching 0.39 tokens per second, Anthropic said. A separate model, Qwen3.5-9B, ran successfully on eight of nine hardware types tested, including Jetson Nanos and Android emulators.
On a closed network of eight emulated Android devices, only one model tested, GLM-5.2, completed a full replication cycle, reaching all eight devices in four of five trials and seven of eight in the fifth, Anthropic said. The company said models with 30 billion parameters or fewer can already complete many of the individual steps required but make mistakes chaining them together end to end.
Anthropic expects smaller models to reach GLM-5.2's current capability "by the end of this year," the company said, if improvement continues at its historical rate. Such malware would spread slowly but would be hard to shut down because it has no central point of control, SecurityWeek reported.
IoT botnets like Mirai have spread for years using simple stolen-credential exploits. An AI model that can reason about unfamiliar devices on its own would not need those credentials, which is a different and harder problem for anyone securing a fleet of smart devices.