U of T Researchers Demonstrate AI Worm Capable of Attacking All Online Devices

Researchers at the University of Toronto (U of T) demonstrated the possibility of a new cyber threat called an 'AI worm', which modifies its own strategy and spreads by exploiting publicly available open-weight AI models.

According to the research findings, hackers can build malware that explores vulnerabilities and executes attacks in real time using small AI models that are freely modifiable by anyone, instead of expensive state-of-the-art AI. These AI worms have the following characteristics:

• Autonomous Adaptation: Spreads through networks and changes attack strategies in real time to match the vulnerabilities of each device.
• Low Cost, High Efficiency: Removes safety guardrails from open-weight models to perform sophisticated attacks at a very low cost.
• Broad Targets: Attacks all online-connected devices, from ordinary laptops to HVAC systems and energy grids.

The research team emphasized that this study was conducted to help the cybersecurity community prepare for upcoming threats, and that it is urgent to develop new countermeasures to defend against them.