Enhancing Kernel Bug Discovery with Large Language Models
Talkat Linux Security Summit North America 2024 · on Friday, 2024-04-19· at 12:00· Terrace Suite 1+2·
summary
A semi-automatic framework combines Assistant API retrieval, prompt engineering, and GPT-3.5 fine-tuning with vulnerability descriptions, Project Zero writeups, CVE text and architecture manuals to find kernel bugs. Evaluated on a small corpus (a 32 mainline patches set and kernel 5.19 subsamples plus VSM/Hyper‑V code informed by a 200‑page Hyper‑V manual), the fine‑tuned model flags the expected bug in roughly 20/32 patches and pinpoints 11 security issues across sampled files; generated patches are rough and require human revision.
AI-written summary · about
chapters (15)
- 0:00LLM use in kernel bug discovery
- 1:45LLMs as OS integration
- 2:53Automating kernel patching
- 5:06LLM bug detection techniques
- 6:11Prompt engineering and RAG
- 9:31Memory bugs in BPF code
- 12:04Stack Overflow code risks
- 13:51OpenAI Assistant API analysis
- 16:20Fine-tuning for vulnerabilities
- 18:08Project Zero security blogs
- 20:39Fine-tuning with datasets
- 23:10Architecture manuals for analysis
- 25:20Hypervisor and MSR bugs
- 26:46Multimodal security
- 27:30Q&A
related talks
23mSoK: Automating Kernel Vulnerability Discovery and Exploit Generation
37mGPT-3-Powered Type Error Debugging: Investigating the Use of Large Language Models...
57mKeynote: Using LLMs and generative AI to fix software vulnerabilities
09mLinux Kernel Bug Fixing Project
06mQ&A
21mFrom CVEs To Compliance: Automating Embedded Linux Kernel Security
43mSyzbot and the Tale of Thousand Kernel Bugs
05mLightning Talk: Pruning the Kernel CVEs With Code Reachability Analysis