Reproduce “RealSec-bench: A Benchmark for Evaluating Secure Code Generation in Real-World Repositories” 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

Yanlin Wang, Ziyao Zhang, Chong Wang, Xinyi Xu, Mingwei Liu, Yong Wang, Jiachi Chen, Zibin Zheng

The paper introduces RealSec-bench, a benchmark for completing Java functions from open-source Maven-based repositories. It evaluates generated code with unit tests and a CodeQL-based, LLM-adjudicated security process. Across 105 tasks, GPT-4.1’s Pass@1 is 15.24%, but SecurePass@1 is 5.71%. Across five models, SecurePass@1 remains below 6%, indicating that test-passing code is not necessarily secure under this evaluation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.