Anthropic Warns of Approaching AI Worm: 15 Cheap TV Boxes Can Now Run a 26-Billion-Parameter Model

Beating AI News Flash: Anthropic has released a study warning that AI could in the future behave like a computer worm, autonomously seeking out internet-connected devices, deploying models, and using new devices to continue spreading. The process would not necessarily require data centers—ordinary connected devices such as smart TV boxes and refrigerators could also become sources of computing power for AI.

Researchers bought 15 Android TV boxes, each priced at about $55, and split a 26-billion-parameter Gemma 4 model to run across these devices. Although it started successfully, inference speed was only 0.39 tokens per second. In another test, Alibaba's Qwen3.5-9B was already able to autonomously set up model inference services on 8 of 9 different types of hardware.

Zhipu's GLM-5.2 stood out in more complex tests. In an isolated simulated network, it completed the process of finding targets, transmitting models, starting inference, and continuing to spread. In 5 tests, 4 covered all 8 target nodes, and the other covered 7. However, the experiment opened an unauthenticated debugging interface in advance, greatly reducing the difficulty of intrusion. GLM-5.2 itself also relied on external servers to run, and what was deployed to the target devices was the smaller Llama 3.1 8B, so it did not achieve true self-replication independent of external computing power.

Researchers also found that AI can use the experience of a previous deployment failure to improve the success rate of the next round. After Qwen3.5-9B passed failure records to the next round of the Agent, the deployment success rate rose from an initial 45% to 61%–84% in subsequent rounds.

Based on the past rate of capability growth of small models, Anthropic predicts that by the end of 2026, small models suitable for running on cheap devices may reach the ability GLM-5.2 currently has to complete simulated spreading tasks.

However, the computing power of cheap hardware remains a limitation. Based on existing experimental data, a 9B model may need one to two days to complete one round of replication.