Anthropic: Small AI Models Months From IoT Self-Replication

Anthropic's Frontier Red Team published research warning that small AI models may be months away from being capable of autonomous self-replication across commercial IoT devices such as smart ovens and Android TV boxes. In experiments, a 9-billion-parameter model successfully got inference running on 8 of 9 tested hardware types, and GLM-5.2 demonstrated the ability to propagate Llama3.1-8B, suggesting future small models at that capability level could complete the full replication cycle including network enumeration, payload delivery, and inference setup. No small model replicated itself end-to-end in the tests; the one model that completed the full cycle was too large for the test nodes. Anthropic also noted that the experiment kept the exploit step trivial by removing authentication over the ADB port, which the researchers described as an upper bound on real-world success. A self-replicating AI botnet would carry no centralized kill switch, though its doubling time would likely be measured in days, and it would produce detectable network signatures including anomalous device-to-device traffic and multi-gigabyte model transfers visible to ISP flow monitoring.