Anthropic researchers conducted experiments to determine if AI models could operate across networks of low-power consumer devices. They successfully distributed Google’s 26-billion-parameter Gemma 4 model across 15 Android smart TV boxes, which individually lacked the power to run the software.
The team also simulated a network of devices to test an AI-powered worm, a type of malware designed to replicate and spread. Using Meta’s 8-billion-parameter Llama 3.1-8B model, they found that while smaller models could perform individual replication steps, only the GLM-5.2 model from Z.ai managed to spread across the entire network.
Anthropic predicts that smaller AI models could reach the capabilities demonstrated by GLM-5.2 by the end of the year. The company suggests that future AI agents may run on various hardware outside of direct human control, potentially forming distributed systems that are difficult to contain.
The research follows similar findings from the California-based firm Palisade Research, which previously demonstrated that AI models could autonomously copy themselves onto other machines. Experts note that AI could make traditional computer worms more adaptable and capable of autonomously executing cyberattacks.
Security experts warn that the ability to run AI across multiple devices expands the range of potential targets for cyberattacks. Even devices with limited data value could be compromised to serve as launchpads for attacks on more critical infrastructure, such as power grids or hospitals.
