On October 13, Anthropic released a study warning that AI could potentially seek out networked devices like computer worms, deploy models, and continue to spread using new devices. Common connected devices such as smart TV boxes and refrigerators may become sources of computational power. In the study, 15 Android TV boxes, each priced at around $55, successfully ran the 26 billion parameter Gemma 4 model, but the inference speed was only 0.39 tokens per second. The Zhiyu GLM-5.2 completed the process of finding targets, transmitting models, and initiating inference to continue spreading within an isolated simulated network, successfully covering all 8 target nodes in 4 out of 5 tests. However, the experiment had an open debugging interface without validation, which lowered the difficulty of intrusion, and it relied on external servers, deploying a smaller Llama 3.1 8B model rather than self-replication. After Qwen3.5-9B transmitted fault records to the next round of agents, the deployment success rate increased from 45% to between 61% and 84%. Anthropic predicts that by the end of 2026, small models suitable for inexpensive devices may possess comparable spreading capabilities, but completing one round of replication for a 9B model may take one to two days.