Anthropic researcher Nicholas Carlini used Claude Code with a simple bash script to discover multiple remotely exploitable Linux kernel vulnerabilities, including a 23-year-old heap buffer overflow in the NFS driver. The script iterated over every kernel source file and prompted Claude Code to find vulnerabilities with no custom tooling. Five kernel vulnerabilities across NFS, io_uring, futex, and ksmbd have been confirmed and patched. The discovery highlights a rapid capability jump in AI-assisted vulnerability research — Claude Opus 4.6 found significantly more bugs than models released just months earlier. Linux kernel maintainers confirm a shift from AI-generated noise to legitimate security reports, with the kernel security list jumping from 2-3 reports per week to 5-10 per day. Security researchers note LLM-based discovery is a new category that combines aspects of fuzzing and static analysis, while raising dual-use concerns about adversaries running the same process at scale.
Sort: