Study warns of self-replicating AI malware using real-time reasoning

Cybersecurity researchers have demonstrated an AI-powered computer worm capable of identifying vulnerabilities, generating attack strategies and spreading autonomously across networks. The study suggests that advances in AI agents could enable a new class of adaptive cyber threats capable of operating with minimal or no direct human intervention.

The research, conducted by teams from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow, describes malware that uses large language models to tailor its behaviour to each target. Unlike traditional worms, the system can adapt its attack methods in real time instead of relying solely on pre-programmed exploits.

Testing in a controlled virtual environment showed the system could successfully compromise multiple machines and replicate across a simulated network over several days. The worm also operated without relying on cloud infrastructure, running AI models locally on infected systems and using those resources to support its operations.

Researchers warned that such capabilities could signal a shift towards what they describe as ‘autonomous generative adversaries’ and stressed the need for stronger detection systems, evaluation frameworks and governance mechanisms. While details were limited to reduce misuse risks, the authors said the findings reflect how rapidly AI-enabled cyber capabilities are evolving.

Why does it matter? 

The research signals a shift in cyber risk from static, signature-based malware to autonomous systems capable of reasoning, adapting, and scaling attacks without human input.

As AI models become more capable and widely deployed, the line between tool and autonomous threat blurs, increasing pressure on cybersecurity systems, patching cycles, and regulation to keep up with real-time, evolving attacks.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our chatbot!