Researchers develop self-replicating AI worm using open-weight models

In 15 runs on a 33-host network, the worm found 31.3 vulnerabilities and compromised 23.1 hosts on average, replicating to 62% of the network over seven days. Unlike traditional worms, it uses an open-weight LLM to generate attack strategies at runtime rather than fixed exploits.

Featured · Nicolas Papernot