Overview
- Researchers published a preprint on June 2 that documents a proof‑of‑concept AI‑driven worm which spread across an isolated 33‑host test network without human intervention.
- The prototype hijacks infected machines’ GPUs to run an open‑weight large language model that reasons about each target and synthesizes tailored exploits at runtime.
- In 15 seven‑day trials the worm on average found about 31.3 vulnerabilities, elevated access on 23.1 hosts, and propagated to 20.4 hosts, while individual exploit attempts succeeded about 44 percent of the time.
- The team withheld key operational details from the public paper and notified Canadian science, security and defence authorities before release to reduce reuse risk.
- Researchers and security experts say immediate steps—prompt patching, zero‑trust and micro‑segmentation, network monitoring, and AI‑assisted testing—are needed because the technique can evade platform controls and will likely improve as models get better.