On 12 October 2026, Anthropic published research warning that advances in small AI models could make autonomous replication across everyday Internet-of-Things (IoT)connected devices a plausible near-term threat. The experiments examined whether AI models could identify potential targets, gain access to devices, establish the computing environment required to run an AI model, and continue propagating without human intervention. The central concern is that an AI-enabled replicator could adapt its deployment to unfamiliar hardware and software, rather than relying exclusively on fixed instructions designed for a specific device. Anthropic did not demonstrate an uncontrolled AI worm spreading across the public internet; instead, it tested individual technical capabilities and their combination under simplified, controlled conditions.
The replication process involves four stages: identifying devices, obtaining the ability to execute code, establishing AI inference and continuing propagation. Inference is the process through which a trained model generates outputs from input. The technical challenge is not simply copying model files to another device, but making the model operational despite differences in hardware, memory, operating systems and software dependencies. An AI agent would need to configure compatible software, resolve installation problems and troubleshoot failures. These tasks could allow it to adapt to devices that were not specifically anticipated by its developers.
by Digital Watch Observatory – Anthropic warns that AI models may be capable of autonomous self-replication within months
