Anthropic tests whether small AI models can copy themselves from device to device

On 12 October, Anthropic published a study on whether small AI models can copy themselves across cheap devices. Models of up to 30 billion parameters complete individual steps but make mistakes when chaining them. A more capable open model, GLM-5.2, did manage it in a simplified test: it spread to 8 of 8 emulated Android devices in 4 of 5 attempts.

Following the trend of open models, the authors estimate that a small one could reach that level by the end of 2026. They themselves warn that this is a rough estimate, that breaking into the devices was trivial in the test, so the result is an upper bound, and that the spread would be slow and could be detected.