作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

智能体驱动的自动化渗透测试:进展、挑战与展望

  • 发布日期:2026-11-18

Automated Penetration Testing Driven by AI Agents: Progress, Challenges, and Prospects

  • Published:2026-11-18

摘要: 渗透测试是评估计算机系统、网络和应用安全性的重要主动防御手段,传统自动化方法较大程度依赖人工经验,在复杂动态环境中难以持续完成环境理解、攻击路径规划、工具调用和漏洞验证等任务。近年来,大语言模型与智能体技术为自动化渗透测试提供了新的技术路径,但现有研究涉及的架构、知识、记忆、推理、工具调用和评测机制相互耦合,相关成果尚缺乏统一的比较和分析。针对上述问题,本研究面向智能体驱动的自动化渗透测试展开调研,主要贡献包括:系统检索并筛选2023年8月至2026年4月期间公开发布或发表的44项代表性研究,从目标阶段、智能体形态、知识增强、记忆管理、规划推理、工具调用、结果验证、评测场景、开源信息等维度对代表性方案进行横向比较。在此基础上,构建面向自动化渗透测试流程的统一分析框架,从多智能体协作架构、领域知识增强、记忆管理、推理规划、工具调用优化及测试评估六个维度归纳关键技术路线,结果表明:多智能体架构正从单一大模型决策演进为角色分工与动态自适应协作模式;检索增强生成与垂类微调可弥补通用大模型在渗透测试领域的知识不足;结构化记忆与上下文接力机制能够缓解长流程测试中的信息衰减问题;工具调用已由单一工具执行发展为多工具协同联动与动态自适应调用;在评测方面,标准化基准的建立推动了不同框架间的横向对比,但现有评测体系在场景真实性、过程性指标覆盖等方面仍存在局限。基于上述分析,本研究进一步凝练了当前研究在长流程任务支撑、复杂任务处理、经验复用、经济成本、评测真实性及安全可控性等方面的关键挑战,并提出面向工具知识深度融合、未知场景适应、持续学习、综合评测和受控自主执行的未来发展方向。本研究可为自动化渗透测试系统的架构设计、技术选型、实验评估和安全部署提供参考。

Abstract: Penetration testing is an important active defense measure for assessing the security of computer systems, networks, and applications. Conventional automated penetration testing approaches rely heavily on human expertise and struggle to consistently accomplish tasks such as environment understanding, attack path planning, tool invocation, and vulnerability verification in complex and dynamic environments. In recent years, large language models (LLMs) and agent technologies have provided new technical pathways for automated penetration testing. However, the architectures, knowledge systems, memory mechanisms, reasoning paradigms, tool invocation schemes, and evaluation and testing frameworks involved in existing studies are highly coupled, and there remains a lack of unified comparison and analysis of relevant research outcomes. To address these issues, this study conducts a systematic investigation into agent-driven automated penetration testing. The main contributions are summarized as follows: This study systematically retrieved and screened 44 representative works first publicly released or published between August 2023 and April 2026, and carried out a horizontal comparative analysis of representative solutions across nine dimensions: target phases, agent paradigms, knowledge augmentation, memory management, planning and reasoning, tool invocation, result verification, evaluation scenarios, and open-source availability. On this basis, this study constructed a unified analytical framework tailored to the automated penetration testing process, and summarized key technical routes from six dimensions: multi-agent collaborative architecture, domain knowledge augmentation, memory management, reasoning and planning, tool invocation optimization, and testing and evaluation. The results indicate that multi-agent architectures have been evolving from centralized single-LLM decision-making to the mode of role-based division of labor and dynamically adaptive collaboration. Retrieval-augmented generation and domain-specific fine-tuning can compensate for the knowledge deficiency of general-purpose LLMs in the penetration testing domain. Structured memory and context relay mechanisms can alleviate the information decay problem in long-horizon testing. Tool invocation has advanced from single-tool execution to multi-tool synergy and dynamically adaptive invocation. In terms of evaluation, the establishment of standardized benchmarks has promoted horizontal comparison among different frameworks; nevertheless, existing evaluation systems still have limitations in scenario realism and coverage of process-oriented metrics. Building on the foregoing analysis, this study further identifies the key challenges of current research in aspects including support for long-horizon tasks, complex task handling, experience reuse, operational cost, evaluation realism, and security and controllability, and proposes future development directions toward deep integration of tools and knowledge, adaptation to unseen scenarios, continual learning, comprehensive evaluation, and controlled autonomous execution. This study can provide a reference for the architectural design, technology selection, experimental evaluation, and secure deployment of automated penetration testing systems.