Researchers from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow have developed a proof-of-concept AI-powered worm that can identify vulnerabilities, devise attack strategies, and spread autonomously across networks. Unlike traditional worms that rely on fixed exploits, this worm uses open-weight large language models running locally on infected machines to adapt its tactics in real time. In tests on an isolated network of 33 devices, the worm identified an average of 31.3 vulnerabilities, compromised 23.1 hosts, and replicated up to seven generations over seven days. It could also exploit vulnerabilities disclosed after its training cutoff by ingesting new security advisories. The researchers withheld some technical details to reduce misuse risk, emphasizing the need for coordinated action to address this emerging threat.

Key facts

• AI worm uses open-weight LLMs running locally on infected machines.
• Identified 31.3 vulnerabilities on average across 33-device test network.
• Compromised 23.1 hosts and replicated up to 7 generations autonomously.
• Can exploit post-training vulnerabilities by ingesting new advisories.
• Researchers withheld technical details to prevent malicious use.

KeyAudit data perspective