Anthropic tests show AI replication across emulated Android devices

Anthropic researchers demonstrated that GLM-5.2, a larger open AI model, could spread across a small network of emulated Android devices in a simplified test that made the hacking step easy. Smaller models could perform many individual steps but still failed to complete the sequence reliably. The team also ran Google's Gemma4-26B across 15 off-brand Android TV boxes bought for about $55 each, achieving less than half a token per second. Anthropic warned that improving small models could eventually spread through poorly secured devices without a central off switch. Large file transfers between devices that normally do not communicate would give internet providers a way to detect the activity.