AI-Powered Worm Compromises 23 Hosts in 7-Day Test, Researchers Show

According to researchers from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow, a new AI-driven worm demonstrated the ability to identify vulnerabilities, generate tailored attack strategies, and spread autonomously across networks. In testing within an isolated virtual environment containing 33 Linux, Windows, and IoT systems, the worm identified an average of 31.3 vulnerabilities, successfully compromised 23.1 hosts, and reached seven generations of self-replication over seven days.

Unlike earlier AI security threats that relied on cloud services, this malware ran large language models directly on infected machines, allowing compromised systems to become part of its computing infrastructure. Researchers noted the worm could exploit vulnerabilities disclosed after the model's training cutoff by ingesting newly published security advisories at runtime, adapting its tactics to different targets in real time.