Toward Intelligent Autonomous Penetration Testing: Integrating AI-Driven Web Vulnerability Discovery with Machine Learning-Based Risk Prioritization
Toward Intelligent Autonomous Penetration Testing: Integrating AI-Driven Web Vulnerability Discovery with Machine Learning-Based Risk Prioritization
Mohammadhossein Daneshpejooh1 Seyed Mohammad Bidoki2
1) B.Sc. in Computer Engineering, Jam Faculty of Engineering, Persian Gulf University
2) Assistant professor in Computer Engineering, Jam Faculty of Engineering, Persian Gulf University
IOI Code
:
AAHD-ABKCKG
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
The expanding attack surface of modern web applications, combined with a persistent shortage of skilled penetration testers, has motivated a decade of research into artificial intelligence (AI) and machine learning (ML) methods for automating security assessment. This paper surveys three converging research threads AI-driven discovery of web vulnerabilities, ML-based prioritization of discovered vulnerabilities, and reinforcement-learning (RL) and largelanguage-model (LLM) driven automated penetration testing and synthesizes them into a single proposed framework, AIVAPT (AIdriven Vulnerability Assessment and Automated Penetration Testing). We review deeplearning classifiers for injection-class vulnerabilities, graph- and sequence-based models for source-level vulnerability identification, and recent LLM-agent systems capable of autonomously exploiting real-world vulnerabilities. We then examine data-driven prioritization approaches that move beyond static severity scores, notably the Exploit Prediction Scoring System (EPSS), and formalize a composite Risk Prioritization Score that fuses CVSS severity, EPSS exploitability, asset criticality, and exposure. Building on this, we cast autonomous exploitation planning as a Markov Decision Process and describe a governance-gated architecture in which a human-in-the-loop authorization layer mediates all autonomous actions. Five comparison tables summarize 30+ peer-reviewed and preprint systems across the three research threads, and we present the governing mathematical formulations (CVSS v3.1, EPSS, standard classification metrics, the Bellman/Qlearning equations, and the proposed composite score). We close with an evaluation protocol grounded in public benchmarks and datasets, and a discussion of open technical challenges and the responsible-use requirements that must govern any such system.
Keywords : penetration testing vulnerability discovery machine learning deep learning reinforcement learning large language models epss cvss web application security autonomous agents