Anthropic Tests How AI Worms Could Spread Across Smart Devices

Anthropic warns that small AI models may be months away from spreading between Internet connected devices without human help, reports Japan Today. The researchers examined whether models could act like computer worms and copy themselves across poorly secured consumer hardware.

In one test, the team divided an open weight Gemma model across 15 inexpensive Android TV boxes. Each box processed only part of the model, allowing the combined network to run it even though no single device had enough computing power.

In a virtual sandbox, a larger model from Z.ai installed a smaller Meta Llama model on target devices. The larger model still depended on a separate server, so the researchers did not demonstrate an autonomous AI worm that could both replicate and run itself on weak hardware.