The introduction of next-generation networks, including 5G/6G, software-defined networking (SDN), and largescale Internet of Things (IoT) deployments, generates huge traffic volumes and dynamic threat surfaces, both of which pose significant challenges to traditional intrusion detection and prevention systems. Machine learning has increased detection accuracy, but most models are opaque, undermining analyst trust and making them challenging to implement in mission-critical settings. This paper proposes a practical intrusion detection and prevention model that combines gradient-boosted decision trees with SHAP explanations and lightweight transformer models, featuring attention-based explainability. Identified intrusions are overlaid on SDN-based mitigation measures, such as flow blocking, rerouting, and rate limiting. A confidence threshold is used to escalate to human analysts when explanations are unclear. Experiments conducted independently on the CICIDS2017 and UNSW-NB15 datasets, along with an emulated SDN testbed that the proposed framework achieves an accuracy of above 94% and generates explanations within less than 5 ms. The detection-toprevention latency of the proposed framework decreases from 12.4 ms to 8.7 ms compared to a baseline IPS. These findings indicate that explainable, real-time IDS/IPS systems can be both highly accurate and transparent, and efficient in prevention, thereby improving trust and accountability in next-generation network defense.