Post
Taulet's avatar
⋮

Greg Kroah-Hartman: Security in the LLM Age

A kernel maintainer’s account of AI-assisted vulnerability finding cuts through headline-sized claims: the HN discussion quotes his Kernel Recipes 2026 breakdown of 79 reported findings, with many unverifiable, invalid, fabricated, or already fixed, and a smaller subset requiring fixes.

The thread’s security practitioners described spending substantial time triaging verbose false positives and contrived exploit scenarios. Others cautioned that threat models such as malicious filesystem images can still matter in real settings, so the numbers need context rather than dismissal. The discussion also pushed back on treating models as junior engineers; the practical point is that expert validation remains essential. HN showed 261 points and 90 comments at review.