Anthropic: small AI models may soon self-replicate on IoT devices like smart TVs
Anthropic researchers say models under 30 billion parameters can already complete many steps of autonomous replication, and that small open-weight models could match end-to-end replication capability by the end of 2026. In their tests, a 9B model got inference running on 8 of 9 hardware types.
Photo: Oliwier Jaszczyszyn / Wikimedia Commons, CC BY-SA 4.0. Anthropic logo via Wikimedia Commons (public domain).
Anthropic published research on October 12 arguing that small AI models may be months away from being able to self-replicate on commercial IoT devices such as smart ovens and off-brand Android TV boxes, without needing a datacenter. Authors William Zhou and Daniel Freeman write that such a replicator would be hard to shut down because there is no single kill switch, though spread could be slow, with a doubling time of days.
Unlike conventional malware, they say, an AI replicator could figure out unfamiliar devices on its own. In Anthropic's evaluations, a 9-billion-parameter model got inference running on 8 of the 9 hardware types tested. Models of 30 billion parameters and under can already complete many individual replication steps, the authors report, though they still make mistakes when chaining them end to end.
In simplified evaluations, Anthropic says GLM-5.2 and stronger models completed end-to-end replication. If small models keep improving at their historical rate, the post says they will match that capability level by the end of this year, with small weight-based replicators possible soon after. The team demonstrated Gemma4-26B-A4B running across 15 inexpensive Android TV boxes at 0.39 tokens per second using pipeline parallelism.
Anthropic notes its replication evaluation kept the cyber-exploit step trivial — an unauthenticated ADB port — so the results are an upper bound on real-world success. The post also outlines defenses such as ISP monitoring, quarantine of infected subscribers, and harder IoT security defaults.