Reproduce “Patch-to-PoC: A Systematic Study of Agentic LLM Systems for Linux Kernel N-Day Reproduction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Juefei Pu, Xingyu Li, Zhengchuan Liang, Jonathan Cox, Yifan Wu, Kareem Shehada, Arrdya Srivastav, Zhiyun Qian
The K-REPRO agentic system, developed at the University of California, Riverside, systematically evaluates LLM-based N-day vulnerability reproduction in the Linux kernel, achieving a 56.0% success rate on 100 KernelCTF bugs. It outperforms state-of-the-art fuzzing methods like SyzDirect by reproducing vulnerabilities in 17.04 minutes on average, compared to 11.74 hours, across a comparable dataset.
Key insights
- LLM agents can achieve non-trivial success in reproducing Linux kernel vulnerabilities (56.0% with XHigh), demonstrating robustness across various kernel subsystems.
- Vulnerability characteristics significantly impact reproduction: race conditions (29.2% success) and temporal memory violations are harder than spatial memory violations, and rich commit messages improve success rates.
- The LLM agent's tool usage involves active static analysis, dynamic verification, and GDB for variable inspection, with minimal impact on success when GDB is removed, suggesting more critical needs in root-cause analysis than dynamic debugging.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.