Autonomous Pentesting using Artificial Intelligence: From the Cybersecurity Point-of-View

• Comparative analysis of traditional and AI-driven penetration testing tools aligned with the MITRE ATT&CK framework. • In-depth evaluation of reinforcement learning-based simulators and emulators for autonomous cybersecurity testing. • Identification of limitations in current RL tools regarding emulation, scalability, and real-world applicability. • Proposal for hybrid simulation-emulation environments to improve realism and transferability in training AI agents. • Novel architecture insights from recent platforms like Cyberwheel and PenGym, advancing autonomous pentesting research. Reinforcement Learning (RL) is emerging as a key approach for autonomous Penetration Testing, enabling agents to learn offensive actions through interaction with simulated networks. However, current RL environments differ widely in realism, attack coverage, and scalability, creating a significant gap between experimental performance and real-world applicability. This work evaluates the internal design of state-of-the-art RL Pentesting simulators, identifying limitations in their action models, MITRE ATT&CK coverage, emulation support, reward mechanisms, network scalability, and environment fidelity. Our contribution is a technical comparison that isolates which environments best support each feature for practical Autonomous Pentesting. Results show that no simulator leads in all area. NASimEmu, Caldera, and Cyberwheel provide the highest environment fidelity; Caldera, Cyberwheel, and CybORG offer the broadest MITRE ATT&CK coverage; and NASim, MulVAL, and CyberBattleSim achieve the best scalability through lightweight graph-based designs.