Researchers Build Self-Replicating AI Worm That Operates Entirely on Local, Open-Weight Models

University of Toronto researchers built and evaluated a proof-of-concept AI-driven worm that runs an open-weight LLM locally to reason about targets, generate tailored exploits at runtime, and self-replicate across a deliberately vulnerable 33-host network; the agent achieved elevated access on ~23 hosts on average and replicated to ~62% of the network over seven days, ingested fresh advisories to exploit recently disclosed CVEs, and demonstrated that GPU-capable victims can serve as distributed inference hubs, complicating containment.