A proof-of-concept AI-powered computer worm developed by researchers at the University of Toronto and collaborating institutions. Unlike traditional malware, which relies on pre-programmed exploits, this experimental worm uses a locally hosted open-weight large language model to analyze targets, devise attack strategies, exploit vulnerabilities, and replicate itself across a network without human intervention. In tests conducted on a controlled 33-host network, the worm autonomously spread to a significant portion of the environment over a seven-day period.
A key innovation is the worm's ability to adapt its behavior in real time. Traditional worms depend on fixed exploit chains that become ineffective once specific vulnerabilities are patched. This AI-driven system instead generates customized attack logic for each machine it encounters, allowing it to target different operating systems, devices, and vulnerabilities dynamically. Researchers found that the worm could identify vulnerabilities, gain elevated access, and replicate itself across both conventional computers and IoT devices without relying on commercial AI services or cloud APIs.
The study raises concerns about a new generation of cyber threats powered by autonomous AI agents. Because the worm runs on open-source models hosted locally, it is not dependent on centralized AI providers that could impose safeguards, usage restrictions, or rate limits. Researchers argue that such systems could create an economic advantage for attackers by allowing compromised machines to provide the computing resources needed for further attacks, reducing the cost of large-scale cyber operations.
The article concludes that while the worm was developed in a controlled research environment to help security experts prepare for future threats, it demonstrates that adaptive, self-sustaining AI-driven malware is no longer purely theoretical. Experts caution that real-world networks are generally more secure than laboratory environments, but the research highlights the need for stronger cybersecurity defenses as AI agents become increasingly capable of reasoning, adapting, and operating autonomously.
