关键词: 大语言模型；智能体；渗透测试；多模态；强化学习；检索增强生成

Abstract: With the accelerated development of artificial intelligence, intelligent agents have demonstrated notable advantages in environmental perception, task planning, and multi-tool coordination. Concurrently, breakthroughs in Large Language Models concerning natural language understanding, logical reasoning, and multimodal processing have provided crucial support for the evolution of intelligent agents. The deep integration of these two technological strands has given rise to LLM-driven autonomous agents for network penetration testing, promoting a gradual shift from the traditional "tool-assisted" paradigm toward "autonomous intelligence." This paper systematically reviews the key challenges and principal technical approaches identified in existing research across four core modules: agent role definition, task planning, memory management, and interactive execution. It further examines the limitations of current methods in areas such as multimodal information processing, automated interaction, and context management. To address these issues, and in view of the ongoing technological evolution of intelligent agents, this paper proposes several promising research directions for intelligent penetration testing. These include multimodal fusion mechanisms, collaborative strategies integrating memory enhancement with reinforcement learning, and knowledge-graph-based vulnerability discovery methods. The analysis indicates that LLM-driven agents for network penetration testing provide substantial technical support for advancing the intelligence and autonomy of cybersecurity operations.

Key words: large language model; agent; penetration testing; multimodal; reinforcement learning; retrieval-augmented generation