Getting LLMs Drunk to Find Remote Linux Kernel OOB Writes (and More) (heyitsas.im)

A groundbreaking vulnerability research team has autonomously discovered over 20 Common Vulnerabilities and Exposures (CVEs), including critical remote, unauthenticated out-of-bounds writes in the Linux kernel’s ksmbd (CVE-2026-31432 and CVE-2026-31433). The innovative approach involves "getting LLMs drunk," leveraging advanced large language models (LLMs) to identify code discrepancies and hunt for vulnerabilities effectively. This method marks a significant shift in vulnerability discovery, as the models can now robustly automate the identification process that traditionally required expert insight. The findings have far-reaching implications for the AI and cybersecurity communities. By automating the vulnerability discovery process, the research not only accelerates the identification of critical security flaws—potentially leading to significant exploits—but also exemplifies the evolving capabilities of AI in practical applications. Noteworthy technical details include how the ksmbd vulnerabilities exploit the kernel's buffer without proper bounds-checking and the harness employed to orchestrate the models’ operation. This work demonstrates the potential for LLMs not only to drive various security tools but to fundamentally alter the landscape of how software vulnerabilities are detected and addressed.