A large language model based penetration test system implementation

Abstract

Penetration testing is a security testing process conducted to identify components and logic flaws in specified information systems that may lead to vulnerabilities, with the aim of preventing these weaknesses from being exploited by malicious actors and strengthening security controls, within the framework of organizational approval and predefined scope boundaries. As corporate networks grow in scale and complexity, conducting these assessments in a regular, repeatable, and scalable manner becomes critical; moreover, due to rapidly changing technology and threat conditions, early detection of vulnerabilities and timely action become even more important. While Large Language Models (LLMs) offer automation potential for certain steps of penetration testing through their natural language processing and reasoning capabilities, the literature has predominantly focused on advanced and commercial models; in contrast, repeatable and measurable evaluation frameworks that reveal under which conditions and to what extent open-source LLMs can produce reliable outputs across different scenarios have not yet sufficiently matured. This thesis presents an initial architecture and evaluation framework for designing a controlled end-to-end autonomous penetration testing workflow using LLM-based planner-coder agents. Within the scope of the study, a FAISS-indexed Retrieval-Augmented Generation (RAG) pipeline is designed using technical PDF sources and structured cybersecurity knowledge sources in order to compile, index, and utilize domain knowledge. In addition, a layered evaluation framework is implemented to compare different open-source planner-coder model variants across reasoning, planning, execution, evidence generation, and reporting dimensions. Rather than focusing on isolated answer quality, the evaluation examines whether the system can maintain technically consistent behavior throughout the autonomous workflow under controlled experimental conditions. In the proposed design, the functions of interpreting technical outputs and making decisions are separated from the generation of scan-oriented commands/scripts, and a role-separated dual-model autonomous architecture is constructed. Experiments are conducted in an isolated laboratory environment using Kali Linux and Metasploitable2; components targeting the automation of stages such as port scanning, analysis, command generation, safe execution, runtime evidence collection, and reporting are designed. At the current stage, it has been observed that, together with RAG-based knowledge support, layered metric evaluation, and output verification approaches, a framework has been established to enable LLM-based structures to produce more consistent assessments over scan outputs and to make the technical basis of system behavior more traceable. The results show that open-source LLM-based planner-coder architectures can support controlled pre-exploitation penetration testing stages such as scan interpretation, planning, safe command generation, evidence collection, and reporting. However, broader validation across more diverse infrastructures, further retrieval integration, and improved autonomous continuity remain future work.