Anthropic shows parts of a self-replicating AI worm on consumer devices

Anthropic’s tests show how such a worm could move across cheap hardware, but also why today’s models still fall short of reliable self-replication.

Anthropic says its red team has shown parts of a self-replicating AI worm that could spread across consumer devices, but the work also highlights why this is not yet a practical threat. In tests, Qwen3.5-9B set up inference on 8 of 9 hardware types, and Gemma4-26B-A4B ran on 15 low-cost Android TV boxes with a combined output rate of 0.39 tokens per second. Anthropic also used eight Google Cloud Android 15 emulators to show network propagation through an open ADB port, with GLM-5.2 completing the sequence across all eight nodes in four of five trials and reaching seven nodes in the fifth. The company says the main blocker is the gap between the model capability needed for reliable replication and the size that can fit on cheap target hardware, though it projects that a 30-billion-parameter model could reach that level by the end of 2026.