Tools
Tools
Ghanem, Mohamed Chahine (2025) Integrating generative LLMs and agentic reinforcement learning for autonomous cybersecurity. In: Hamilton Institute Seminar, 01/10/2025, Maynooth, Ireland.
Generative large‑language models (LLMs) and agentic AI underpinned by reinforcement learning (RL) are transforming the cybersecurity research landscape. This seminar abstract outlines three interconnected strands of enquiry; First, we consider generative LLMs for threat modelling, demonstrating how transformer‑based architectures can synthesise polymorphic malware pseudocode, construct realistic attack narratives and produce high‑fidelity phishing exemplars. Incorporation of LLM‑augmented datasets into anomaly detection pipelines has been shown to enhance adversarial robustness and reduce false‑positive rates in benchmark evaluations. Second, we explore agentic AI orchestration, wherein autonomous agents employ RL‑driven planning techniques to observe network telemetry, devise multi‑step mitigation strategies and execute adaptive countermeasures within emulated cyber‑range environments. We assess performance trade‑offs between latency and throughput, and introduce methods for embedding explainability to preserve human auditability. Third, we survey deep RL approaches to dynamic policy optimisation, including proximal policy optimisation and multi‑agent reinforcement learning frameworks applied to software‑defined networks. Case studies illustrate how RL agents iteratively refine system configurations to withstand evolving adversarial tactics. Finally, we identify critical research challenges: certifying automated defences, aligning generative outputs with security objectives, and establishing interdisciplinary governance and ethical oversight. Attendees will acquire a comprehensive roadmap to advance AI‑powered cybersecurity innovation.
Download (7MB) | Preview
 |
View Item |