Anthropic flags AI risk of spreading across cheap home devices

Anthropic published research Monday indicating that small AI models may be only months away from spreading across poorly secured internet-connected devices without human help. The San Francisco AI company tested whether models could locate vulnerable hardware, copy themselves onto it and keep propagating. It found models under 30 billion parameters could complete many of the required steps, though they still struggled to execute the full chain consistently. In simplified tests with some security barriers removed, GLM-5.2 and stronger systems managed end-to-end replication. Anthropic said gadgets such as off-brand Android TV boxes, smart ovens and other IoT devices are attractive targets because they often have weak defenses and no central shutoff point.