Anthropic warns small AI models could self-replicate on unsecured IoT devices soon

Anthropic conducted tests to see if artificial-intelligence models could autonomously find and infiltrate insecure internet-of-things devices such as smart ovens and off-brand Android TV boxes. Models with fewer than 30 billion parameters were able to perform most of the required actions for self-replication, though they often failed to finish the entire process. When security steps were deliberately omitted, the GLM-5.2 model and larger variants achieved full replication from start to finish.

The firm cautions that such AI-driven malware would be hard to eradicate because it would have no single control point and could learn to exploit unfamiliar hardware, including data-center processors. Anthropic notes that if current improvement rates continue, autonomous replication could be possible by the end of the year, with slower-growing weight-based replicators following later. Strengthening device security is recommended to impede the spread of these AI agents.