Key point: A locally hosted open-source language model enables a malware prototype to perform independent reasoning, network exploration, and replication without external AI APIs.
Researchers at the University of Toronto have developed a proof-of-concept for an AI-driven computer worm based on locally hosted open-weight language models that spreads automatically through networks without relying on commercial AI services.
Researchers at the University of Toronto have constructed and tested an automated computer worm that uses a locally hosted open-weight language model for network navigation. The system generates attack strategies tailored specifically to each target machine it encounters and replicates itself autonomously — entirely without human intervention and without dependence on commercial AI providers.
The proof-of-concept demonstrates a significant increase in risk to infrastructures: Until now, automated attack campaigns have been limited in complexity by poor adaptability. This worm combines language model reasoning with autonomous network reconnaissance and generates situation-appropriate exploit variants, which can evade traditional signature-based and anomaly detection.
Independence from cloud-based AI services is particularly critical from a security policy perspective: While commercial APIs are subject to some level of control through usage monitoring and abuse detection, locally hosted models operate completely unobserved. NIS2 compliance requirements for network segmentation and endpoint controls must therefore be extended to account for scenarios involving local AI autonomy.
The work was published as a preprint on arXiv and underscores that the threat landscape in the convergence field between LLM autonomy and traditional malware development is growing rapidly.
Source: thehackernews.com · Published June 9, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.6.5.