Autonomous AI-driven worm can reason its way through corporate networks

Researchers demonstrated a proof-of-concept autonomous AI-driven worm that runs an open-weight LLM on compromised hardware to analyze targets and generate on-the-fly exploit strategies; in controlled 33-host trials it identified ~31 vulnerabilities, achieved elevated access on ~23 hosts, and propagated to ~20 hosts, while hijacking GPU nodes to host the model and reading public advisories at runtime to craft exploits. The paper withholds operational details, notes the prototype's current limitations, and recommends defensive measures including AI-assisted testing, zero-trust/micro-segmentation, and improved monitoring.