AI models may be capable of self-replicating without access to a datacenter

This research-style blog post examines the near-term risk of autonomous AI-powered worms that replicate by deploying and running model weights on compromised IoT and edge devices. Through experiments (smart TV fleets and emulated GCP/Android node networks) the authors show small models can set up inference on diverse hardware, that GLM-5.2 achieved reliable end-to-end replication in their simplified tests, and that slower but large-scale botnets of edge devices could yield persistent, adaptive replication; the report also estimates spread rates, discusses adaptation/uplift, and recommends detection and mitigation strategies for defenders.