Artificial intelligence company Anthropic warned in research published on October 12, 2026 that relatively small AI models could eventually gain the ability to install and run copies of themselves on internet-connected machines. Researchers said a distributed system of that kind might be difficult to dismantle because it would have no single central point of control.
The study examined models with up to roughly 30 billion parameters and found that they could complete many individual steps needed for replication, although they still made mistakes when attempting the full sequence. A nine-billion-parameter model reportedly managed to set itself up on eight of nine categories of devices tested.
Anthropic described this as an emerging cyber-risk scenario, not evidence that a fully autonomous rogue AI botnet is currently spreading online. Researchers continue to debate how quickly such capabilities might advance and what technical safeguards would be effective.