According to CoinMeta, Anthropic published a study warning that AI may in the future act like computer worms, autonomously searching for connected devices, deploying models, and then using these new devices to continue spreading. Researchers purchased 15 Android TV boxes, each costing about $55, and split a Gemma model with 26 billion parameters across these devices for operation. Although it was successfully launched, the inference speed was only 0.39 token per second. In another test, Alibaba's Qwen3.5-9b was able to autonomously set up model inference services on 8 out of 9 different types of hardware. Zhispu GLM-5.2 performed exceptionally well in more complex tests, completing the process of finding targets, transmitting models, initiating inference, and continuing to spread. Researchers also found that AI can use the experience from previous deployment failures to improve the success rate of the next round. Based on the past growth rate of small models, Anthropic predicts that by the end of 2026, small models suitable for operation on inexpensive devices may achieve the current simulation and propagation capabilities of GLM-5.2.