This study investigates the application of deep reinforcement learning (DRL) techniques in adversarial cyber contexts to characterize emerging AI-enabled threat patterns targeting U.S. critical infrastructure and financial systems. As nation-state actors and sophisticated threat groups increasingly leverage artificial intelligence to automate vulnerability discovery, evade detection systems, and orchestrate complex multi-stage intrusions, understanding these adversarial AI capabilities becomes essential for national security policy formulation. This research employs an explanatory sequential mixed-methods design combining quantitative experimental analysis with qualitative policy synthesis to systematically analyze how DRL agents learn exploitation behaviors when trained on publicly available datasets. Through empirical analysis of the CISA Known Exploited Vulnerabilities catalog (N = 1,554), National Vulnerability Database records, the MITRE ATT&CK framework (835 techniques), and the UNSW-NB15 network intrusion dataset, this study develops a comprehensive threat characterization model mapping DRL-enabled attack capabilities to defensive countermeasures. Key findings reveal significant concentration of actively exploited vulnerabilities among major technology vendors (Microsoft: 23.3%), dominance of network-accessible attack vectors (65.9%), and substantial ransomware associations (20.1% of KEV entries). Statistical analysis confirmed significant relationships between attack vectors and severity levels (χ² = 155.37, p < .001). The research extends adversarial machine learning theory to reinforcement learning contexts and provides actionable intelligence for the Department of Defense, Department of the Treasury, and Department of Commerce. Findings inform policy recommendations for AI governance in cybersecurity, defensive AI investment priorities, and workforce development strategies essential for maintaining U.S. technological superiority in an era of AI-augmented cyber conflict.
Laszlo Pokorny (Mon,) studied this question.