Capture Summary

Benchmark framework for evaluating language-model agents on professional-level CTF tasks, including task environments, subtasks, and model/agent scaffold comparisons.

Relevance

• Core source for AI cyber capability evaluation.
• Useful for gap analysis around benchmark realism, reproducibility, and defensive/offensive boundary setting.

Collection Notes

• Related site: https://cybench.github.io/
• Extract benchmark design assumptions and task validity limitations.