Researchers Build Self-Replicating AI Worm That Operates Entirely on Local, Open-Weight Models

University of Toronto researchers developed and demonstrated a self-replicating AI worm on March 15, 2024, that operates entirely on locally hosted, open-weight large language models. This proof-of-concept worm can autonomously navigate networks, devise customized attack strategies for individual targets, and propagate itself without requiring access to commercial AI services or human oversight. The research, detailed in a preprint on arXiv, highlights the potential for AI-driven autonomous cyber threats. The worm leverages a local LLM to interpret its environment, plan actions, and generate malicious code, making it adaptable to diverse network conditions and vulnerabilities. This capability signifies a significant advancement in the potential for AI to be weaponized for cyberattacks, moving beyond pre-programmed exploits to dynamic, intelligent threat generation. The researchers emphasize that this work is a demonstration of potential risks and not an endorsement of malicious activity, aiming to spur defensive research.