HN – Show HN: I benchmarked LLM agents on fixing real-world security vulnerabilities

I built a benchmark with 20 real CVEs across 18 Python projects (Pillow, GitPython, yt-dlp, urllib3, etc). I've run it over 5 LLM agents (3 OpenAI, 2 poolside) and 3 different prompts (full advisory, locate, diagnose) with a total of 300 runs. The agents are tasked to fix security vulnerabilities in a sandboxed environment and they are scored against a hidden security tests from the maintainer's own fix.

Best solve rate was 50%. On the other 50%, some fixes are sometimes coherent and pass all regression tests, but vulnerability still present.

The main differentiator I found between models is cost: gpt-5.5 at 12× more expensive than gpt-5.4-mini while producing statistically similar results. Within-family performance gaps are small, which points out the difference is likely due to model training data. I also did a power analysis and the task count needed to detect a meaningful within-family edge at ~700.

Full write-up: https://giovannigatti.github.io/cve-bench

Code: https://github.com/GiovanniGatti/cve-bench

Show HN: I benchmarked LLM agents on fixing real-world security vulnerabilities

1 comments