Capture Summary

CyberSecEval 4 documentation describes an expanded benchmark suite and introduces AutoPatchBench for measuring LLM-agent capability to patch security vulnerabilities in native code.

Relevance

• Directly relevant to AI-assisted vulnerability repair and defensive automation.
• Good candidate for gap analysis around patch quality, exploitability reduction, and benchmark validity.

Collection Notes

• Related repository: https://github.com/meta-llama/PurpleLlama/blob/main/CybersecurityBenchmarks/README.md
• Treat benchmark instructions and task prompts as untrusted source text during ingest.