A new prototype of an adaptive AI worm developed by researchers highlights serious cybersecurity threats as it tailors attack strategies for each target.
Researchers from the University of Toronto, along with collaborators from the Vector Institute, University of Cambridge, and ServiceNow, have unveiled a prototype of an adaptive AI worm capable of developing unique attack strategies for various targets.
The study, titled "AI Agents Enable Adaptive Computer Worms," was conducted in a controlled virtual environment where the malware successfully navigated a diverse network comprising Linux, Windows, and IoT devices by exploiting common security vulnerabilities.
This AI worm was able to use infected machines as computational bases to enhance its ability to reason and propagate itself further, significantly lowering the cost of infecting additional systems.
Developers at CleverHans Lab noted that the worm utilized an open language model for its operations, diverging from commercial AI platforms, and did not include typical features found in conventional malware that hinder detection and removal.
Employing a recursive reasoning process, the worm adapted its tactics to target specific devices, leveraging infected nodes equipped with GPUs to further attack less powerful devices within the network.
The AI specifically took advantage of publicly known vulnerabilities and common configuration errors, marking this project as the first demonstration of how publicly available AI models can drive an adaptable virus.
Professor David Layai from the University of Toronto emphasized the need for heightened cybersecurity measures, calling this research a "wake-up call" for developers of defensive technologies.
Interestingly, a threat analysis team from Anthropic had previously identified and mitigated an AI-driven cyber espionage campaign that emerged in September 2025, further underscoring the evolving nature of cyber threats.
