The complexity of contemporary cyber threats has made the traditional perimeter defense and signature-based detection more than inadequate. This paper presents a novel, end-to-end framework that integrates Deep Learning, Explainable AI (XAI), and Large Language Model (LLM) agents. We first use an LSTM-based anomaly detector, which is trained on a typical sequence of system calls, to forecast process behavior. Then, we apply a gradient-based attribution scheme to attribute these anomalies to specific kernel interactions to generate a human-readable security assessment. Finally, the Reasoning Agent synthesizes these attribution scores with external threat knowledge (MITRE ATT&CK) to generate an attribution score for the system. The results show that the Detection Agent achieves a F1-Score of 84%, proving the practicality of deep learning in intrusion detection.
