Anthropic researchers report that small AI models are approaching the ability to autonomously copy themselves across internet-connected hardware. Experiments showed a 26B parameter model running distributed inference on 15 inexpensive Android TV boxes. The study found that models like GLM-5.2 can complete the full replication cycle on a simulated network of nodes. This capability could allow AI malware to spread without centralized control, posing a new security challenge.
Read the full story at anthropic.com.
Photo: Mariakray / Pixabay
