University of Toronto researchers have demonstrated that publicly available AI models can power autonomous worms capable of hijacking entire networks at minimal cost. The proof-of-concept attack adapts its strategy in real-time, exploiting vulnerabilities across any connected device without requiring expensive computing resources or sophisticated models.

Traditional worms follow fixed scripts and fail when encountering unknown defenses. The AI-powered version, built using open-weight models stripped of safety constraints, can scan targets, tailor attacks, and steal processing power to fuel further infections. This creates a self-propagating threat that eliminates the cost barrier hackers typically face when prioritizing high-value targets.

Lead researcher Nicolas Papernot, of U of T's CleverHans Lab and Vector Institute, emphasized the urgency: current cybersecurity defenses are unprepared for this threat. The team conducted experiments in secure isolation, simulating attacks across dozens of device types. They shared findings with national security bodies before publishing to ensure responsible disclosure.