Researchers Demonstrate Self-Replicating AI Worms Using Open-Weight LLMs

Researchers from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow have built a proof-of-concept computer virus that demonstrates a new threat: self-sustaining AI worms that use compromised GPU resources to run open-weight large language models for reasoning and attack planning. The worm uses stolen computing power to generate tailored attack strategies, detect vulnerabilities, and execute exploits against new targets to spread itself. In testing, the system achieved an 80% success rate at vulnerability detection, 53% at exploitation, and 88% at self-replication, for an overall success rate of approximately 37%. The researchers warn that autonomous generative adversaries powered by AI are no longer theoretical threats and organizations must prepare for such threats.