Researchers Build Self-Evolving AI Worm Using Open-Source Large Language Models

Research teams at the University of Toronto and the University of Cambridge used free, open-source large language models to build an AI worm that can do more than replicate itself. The malware can analyze vulnerabilities in connected devices and generate its own methods of attack. Such automated, adaptive attacks could weaken traditional cybersecurity defenses that rely on fixed signatures and post-attack patching.

The team's latest demonstration showed that the worm could switch targets in real time as it spread and dynamically adjust its attack strategy for different victim devices, instead of relying on a single preprogrammed exploit. The researchers warned that patching known vulnerabilities may therefore fail to stop subsequent mutations of an attack. The accompanying materials did not specify a publication date or transaction value, and said only that the models were available free of charge.