Agentic AI-Driven Cyber Defense Framework for Autonomous Threat Detection and Response

Abstract

The proliferation of cloud computing, Internet of Things (IoT), edge computing, and distributed enterprise networks has made the security landscape much more complex than ever before and has created a need for traditional rule-based security measures to be incapable of detecting advanced and ever-changing cyber threats. This research proposes a framework for Autonomous Threat Detection and Response (ATDR) based on Agentic AI which integrates multiple agents of Artificial Intelligence (AI), Transformer networks, Graph Neural Networks (GNNs), Reinforcement Learning (RL), Security Knowledge Graphs (SKGs), and Explainable Artificial Intelligence (XAI) to facilitate intelligent threat detection, context-aware reasoning, adaptive decision-making and autonomous incident response. The dataset CICIDS2017 was used to test the framework and it was implemented using the GPU acceleration of NVIDIA RTX 4090 on a Python environment. The accuracy, precision, recall, F1, AUC and response time of experimental results are 99.36%, 99.08%, 98.94%, 99.01%, 99.81% and 63ms, respectively, which are better than five recent state-of-the-art methods. The improvements were statistically significant $(\mathbf{p}<\mathbf{0. 0 0 1})$ according to the statistical analysis (ANOVA). The proposed framework is scalable, explainable and adaptive for the security of modern cyber infrastructures for advanced and evolving attacks.