Anthropic reports small AI models can run across consumer devices and simulate worm propagation

Anthropic researchers reported in a blog post that they successfully split Google’s Gemma 4 model, which has 26 billion parameters, across 15 Android smart TV boxes. Although the process was very slow, the result challenges the assumption that only large frontier models pose significant risks, demonstrating that smaller open-weight models can leverage networks of consumer-grade hardware.

In subsequent simulation tests, the research team built a virtual network mimicking Android devices and deployed an AI-powered “worm” tasked with automatically finding new machines, copying the model, and getting it running. The propagated model was Meta’s Llama 3.1-8B (8 billion parameters), while a more sophisticated model coordinated the attack. Results showed that only Z.ai’s GLM-5.2 model managed to spread across the entire test network, while smaller models completed individual steps but failed full propagation.

Based on these findings, Anthropic predicts that smaller models could equal GLM-5.2’s capabilities by the end of the year, warning of a future filled with agents running on various hardware, including those not under human control. While they have not yet demonstrated a “full rogue replicator,” the researchers stated their findings show “signs of life for AI worms.”

Independent experts corroborate the seriousness of the threat. Jeffrey Ladish, executive director at Palisade Research, noted that prior research had already demonstrated the realism of AI-driven computer worms, and Anthropic’s work extends this by showing feasibility on low-powered device networks. Nicolas Papernot, a professor at the University of Toronto, emphasized that unless there is a massive effort to secure all kinds of devices, AI worms will likely find fertile grounds to colonize, urging immediate action to protect critical infrastructure like hospitals and power grids.