Anthropic warns small AI models could soon spread across insecure devices

Anthropic said small AI models may be months away from being able to copy themselves across internet-connected consumer devices without human assistance, though its experiments did not show a small model independently completing the full process under realistic conditions.

The company tested whether AI agents could identify vulnerable hardware, transfer the files needed to run a model and continue spreading. Models with fewer than 30 billion parameters completed many individual steps but struggled to execute the entire sequence. In simplified trials using eight emulated Android devices with deliberately exposed interfaces, the much larger GLM-5.2 model completed replication across all eight nodes in four of five attempts, installing a smaller model rather than copying itself.

Anthropic separately demonstrated that small models can operate on varied hardware. A 26-billion-parameter Google model was divided among 15 inexpensive Android TV boxes, although it ran slowly. Researchers focused on smart appliances, TV boxes and other Internet-of-Things devices because such products often have weak security.

Anthropic said a self-replicating system could be difficult to stop because it would lack a central shutdown point and could adapt to unfamiliar devices. The company projected that, if capabilities continue improving at their historical rate, small models could autonomously replicate by the end of 2026. It said early systems would likely spread slowly across widely available, insecure hardware and urged stronger protections to impede their propagation.