Anthropic Tests Show AI Replication Steps, Not Real-World Spread

Anthropic said its controlled tests found that some AI models could complete the steps needed to replicate across a simplified network, while warning that smaller models could develop the capability if they continue improving. The company stressed that the experiments did not demonstrate autonomous spread through commercial data centers or establish an imminent real-world threat. The findings, published Monday, describe a possible route to AI-powered worms that could move among internet-connected devices such as cameras and household appliances. Such networks could be difficult to contain if they spread without a central point that defenders could disable, Anthropic said. The company’s report focused on how increasingly capable AI agents might use the computing power of connected devices to run other models. Anthropic said models with 30 billion parameters or fewer could already perform many individual tasks involved in self-replication, but struggled to carry out the full sequence reliably. In tests using a simplified network with some security steps removed, GLM-5.2 completed end-to-end replication, according to the company.