Leveraging Large Language Models for Challenge Solving in Capture-the-Flag

Capture-the-Flag (CTF) competitions are a prominent method in cybersecurity for practical attack and defense exercises. Despite the rapid advancements in Large Language Models (LLMs), their potential for solving CTF challenges remains underexplored. In this paper, we first propose a flexible CTF platform designed to reflect real-world penetration testing scenarios, bridging the gap between theoretical learning and practical cybersecurity challenges. Our platform is highly customizable, freely deployable, and capable of generating scenarios that closely mirror real network vulnerabilities. More importantly, we introduce an automated LLM agent framework that tackles CTF challenges using an integrated toolchain and various plugins to enhance problem-solving efficiency. In addition, we propose a human-validated LLM agent framework to address the potential limitations of the fully automated LLM agent, providing a clearer evaluation of the LLMs' intrinsic capabilities. We evaluate the agent's performance using four LLMs: GPT-4o, GPT-4o mini, o1-preview, and o1-mini. Although the LLMs' performance on dynamic and complex penetration tests reveals certain limitations, which need further exploration, our experimental results demonstrate that LLMs can leverage their extensive knowledge bases to effectively solve CTF challenges.