Self-Replicating AI Worm Spreads to 20 Machines in 7 Days

Researchers demonstrate an AI-driven worm that autonomously exploits vulnerabilities and replicates across networks, achieving high infection rates in controlled lab tests.

Woofun AI reports that a consortium including the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow developed a computer worm powered by an open-weight large model. The system autonomously scans networks, identifies vulnerabilities, generates attack plans, and replicates upon compromising devices. In 15 rounds of experiments within an isolated network of 33 virtual machines, the worm gained administrative access to an average of 23.1 machines and spawned replicas on 20.

4 machines over seven days, reaching seven generations of propagation. Unlike traditional worms using pre-set scripts, this system adjusts strategies based on individual machine contexts and translates text-based vulnerability data into executable attack steps. The study remains a laboratory proof-of-concept; test machines lacked antivirus software or active defenses, relied on preloaded vulnerabilities, and utilized a shared GPU pool, with the model and code not yet public.