[Anthropic Warns of AI Worm Risks: 15 Low-Cost TV Boxes Can Run a 26 Billion Parameter Model] According to a report by Jinse Finance, on October 13, Anthropic released a study warning that AI in the future might behave like computer worms, autonomously seeking networked devices, deploying models, and using new devices to continue spreading. Common networked devices such as smart TV boxes and refrigerators could become sources of computational power. In the study, 15 Android TV boxes priced at approximately $55 each successfully ran the 26 billion parameter Gemma 4 model, though the inference speed was only 0.39 tokens per second. The Zhipu GLM-5.2 model completed the process of locating targets, transmitting the model, initiating inference, and continuing propagation within an isolated simulated network, covering all 8 target nodes in 4 out of 5 tests. However, the experiment preemptively opened a verification-free debugging interface, reducing the difficulty of intrusion, and relied on external servers to deploy a smaller Llama 3.1 8B model rather than achieving autonomous replication. After Qwen3.5-9B transmitted fault logs to the next round of agents, the deployment success rate increased from 45% to 61%–84%. Anthropic predicts that, given the rapid improvement in the capabilities of small models, by the end of 2026, small models suitable for low-cost devices may possess equivalent propagation capabilities. However, it may take one to two days for a 9B model to complete a single replication cycle.