On October 13th, Anthropic's team responsible for testing cutting-edge AI risks released research stating that small, open-source models may be able to autonomously replicate on inexpensive IoT devices such as smart TV boxes within a few months, forming "AI worms" that don't require data centers. Researchers connected 15 generic Android TV boxes (around $55 each) and successfully ran the Gemma4-26B model in segments.
Experiments showed that the 9-byte Qwen3.5 could autonomously build inference environments on 8 out of 9 types of unfamiliar hardware. In an 8-node simulated network with simplified intrusion steps, GLM-5.2 reliably completed the full cycle of "scanning the target, copying itself, starting inference, and continuing to spread," infecting all nodes in 4 out of 5 trials. Based on the historical growth rate of small model capabilities, models below 30 bytes are expected to reach this level by the end of 2026.
The research also found that the replicator accumulates experience through intergenerational note-taking, increasing the success rate of setup from 45% to a maximum of 84%. However, this type of worm spreads slowly, with a single replication taking about one to two days. Even at a scale of 2 million devices, the computing power is only equivalent to about 40 H100 processors. The author suggests that the spread can be contained through ISP traffic monitoring and isolation, and by removing default passwords from devices. The author emphasizes that this is currently only a "sign of life" and not a ready-made threat.
