Anthropic published research on 12 October 2026 arguing that small AI models may be months away from self-replicating on commercial IoT devices such as off-brand Android TV boxes and other weakly defended home gadgets. The work is a controlled evaluation, not a report of an active worm in the wild.
What they actually demonstrated
The team ran Gemma4-26B-A4B across 15 inexpensive Android TV boxes (about $55 each) with pipeline parallelism, at roughly 0.39 tokens per second. Separately, they tested whether open-weight models can set up inference on unfamiliar hardware and chain a simplified replication loop (scan, open an unauthenticated ADB path, copy a payload, stand up inference).
A 9B-class model got inference running on 8 of 9 hardware types they tried. Models at 30B and under already complete many individual steps but still make mistakes when chaining them. In their simplified end-to-end eval, only GLM-5.2 completed reliable replication across a firewalled fleet of Android emulators; the cyber step was deliberately easy (open ADB without authentication). Anthropic treats that as an upper bound, while noting that past IoT botnets such as Mirai spread with default passwords.
They estimate that if small models keep improving at the recent Epoch Capabilities Index trend, a model small enough to live on 4 GB TV boxes could reach GLM-5.2-class replication ability by the end of 2026. Spread on that class of hardware would be slow - on the order of days per generation in their timing tables - more like a persistent "digital slime mold" than a fast datacenter worm.
Why smart-home operators should care
Home and small-business IoT is the near-term target in their threat model: lots of devices online, historically weak defaults, and enough RAM when sharded to run sub-30B models. Devices with built-in AI accelerators (some cameras, conferencing gear) raise the ceiling further. Unlike classic malware tied to one firmware family, a weight-based replicator could adapt notes and code as it hops hardware.
Defences they highlight: watch ISP flow data for multi-gigabyte model transfers and unusual device-to-device inference traffic; quarantine infected subscribers the way operators cleaned Mirai; and push manufacturers toward secure defaults (the UK already bans default IoT passwords). For a Home Assistant household the practical takeaway is familiar: isolate IoT VLANs, change defaults, keep firmware updated, and do not expose management ports such as ADB to the LAN or internet.