Getting LLMs Drunk to Find Remote Linux Kernel OOB Writes (and More)
Summary
An in-depth look at using autonomous LLM-based agents to hunt vulnerability bugs across Linux kernel ksmbd and other software. The article outlines a multi-agent harness, highlights CVEs discovered, and discusses architecture, drunk prompting experiments, and future research directions. It emphasizes that orchestration and external review can unlock vulnerability research with smaller models.