Penetration testing plays a vital role in identifying security weaknesses in modern computing systems. With the rapid growth of distributed architectures, cloud-native applications, and microservices, traditional penetration testing approaches have become increasingly complex and time-consuming. Although automated tools are widely used, they typically function in isolation and require significant human expertise to coordinate multi-stage attack scenarios This paper presents an enhanced AI-driven penetration testing framework that leverages the Model Context Protocol (MCP) for structured communication and orchestration among multiple intelligent agents. The proposed system integrates reconnaissance, vulnerability assessment, exploitation, privilege escalation, and reporting into a cohesive pipeline. Unlike traditional systems, the framework incorporates contextual reasoning, adaptive decision-making, and dynamic exploit chaining using an AI Planner. Additionally, the system constructs real-time attack graphs and computes risk scores based on vulnerability severity, exploit confidence, and attack depth. Experimental results demonstrate significant improvements in automation efficiency, reduction in manual effort, and higher success rates in identifying complex exploit chains. The proposed framework represents a shift from static automation toward intelligent, adaptive penetration testing systems.
M et al. (2026) studied this question.