urllib3

A benchmark tested 5 LLM agents on fixing 20 real-world security vulnerabilities across 18 Python projects. The best solve rate was 50%, with cost differences between models (e.g., gpt-5.5 vs. gpt-5.4-mini) outweighing performance gains, likely due to training data variations.