Researchers demonstrate an AI worm that adapts as it spreads

Researchers from the University of Toronto, Vector Institute, and University of Cambridge demonstrated a proof-of-concept computer worm powered by publicly accessible AI models. Unlike traditional worms that repeatedly use one exploit, the system reasons about each target, selects or creates an attack approach, and then propagates. In a simulated corporate network, reporting says it compromised a large majority of targets within a week without human involvement. The work was conducted in a secure lab, disclosed to authorities, and published as a preprint rather than a peer-reviewed result. It nevertheless exposes a difficult security shift: patching one vulnerability or relying on centralized model-provider safeguards may not stop adaptive malware built with freely available models. Defenders will need stronger isolation, monitoring, segmentation, and behavior-based detection.

Key details: June 2, 2026, University of Toronto, Vector Institute, University of Cambridge, Publicly accessible AI models, Adaptive self-propagation, Controlled lab study.

Continue swiping for more AI Brief stories.