Anthropic Research Moves AI Self-Replication Risk Beyond Data Centres

San Francisco, 13 October 2026 – New research from Anthropic suggests that small artificial-intelligence models may be months away from possessing enough capability to replicate across commercial internet-connected devices without relying on a conventional data centre. The finding does not show that a fully autonomous rogue system exists, but it narrows the gap between a theoretical risk and an operational cybersecurity problem.

Researchers tested whether models could complete key stages of replication: finding targets, gaining execution, establishing inference and continuing propagation. Models with 30 billion parameters or fewer can already perform many individual steps, although most still make mistakes when combining the entire sequence. In simplified evaluations, GLM-5.2 completed end-to-end replication reliably.