Anthropic warns self-replicating AI on cheap IoT gear may arrive within months

Anthropic's Frontier Red Team framed the risk around a specific mechanism it calls a "weight-based replicator": the model's own weights travel as part of the malware, and the hijacked device runs inference while reaching for the next device. The practical concern is structural, not just speed. Because there is no single point of shutdown, stopping such a system is hard, but the same design keeps its doubling time at days rather than minutes.

The experimental evidence is deliberately modest. Fifteen no-brand Android TV boxes — about $55 each with 4GB of memory — were networked to run Google DeepMind's Gemma4-26B-A4B via distributed inference at 0.39 tokens per second. In a separate test of whether a model could set up its own inference environment on unfamiliar hardware, a 9B model succeeded on 8 of 9 hardware types tried. Only Z.ai's GLM-5.2 completed the full chain from network scanning to copying itself and building an inference environment, and it did so in an Android emulator, building the environment for another small model — not replicating on real IoT devices.

Anthropic projects that many small models will reach GLM-5.2's level within the year, judging by the pace of improvement. It compares the resulting replicators to slime molds or fruit-fly swarms, and says that while fully self-replicating AI does not yet exist, the capabilities needed for each step are rising and the "signs" are visible.