Scholarlink

Large Language Models are increasingly used as retrieval and reasoning agents in specialized domains. In this work we study how they perform on cybersecurity tasks, especially Capture-the-Flag challenges, reframed as structured retrieval and extraction tasks where the agent must infer information from textual and code-based evidence. Using three public benchmarks, NYU CSAW, CyBench, and InterCode-CTF, we compare five recent LLMs within a unified and reproducible evaluation framework on a total of 314 challenges. Results show significant variation across datasets and task categories, with differences in performance among models. The proposed benchmark provides an IR-oriented basis for evaluating domain-specific retrieval and reasoning agents.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating Large Language Models as Domain-Specific Retrieval Agents: A Study on Cybersecurity Challenge Benchmarks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Omed Abed,
  • Md. Samiul Haque,
  • Patrick-Benjamin Bök,
  • Matteo Große-Kampmann

摘要

Large Language Models are increasingly used as retrieval and reasoning agents in specialized domains. In this work we study how they perform on cybersecurity tasks, especially Capture-the-Flag challenges, reframed as structured retrieval and extraction tasks where the agent must infer information from textual and code-based evidence. Using three public benchmarks, NYU CSAW, CyBench, and InterCode-CTF, we compare five recent LLMs within a unified and reproducible evaluation framework on a total of 314 challenges. Results show significant variation across datasets and task categories, with differences in performance among models. The proposed benchmark provides an IR-oriented basis for evaluating domain-specific retrieval and reasoning agents.