Automated analysis of surveillance video plays a critical role in modern security and public safety systems. In weapon-related activity recognition, high classification accuracy alone is insufficient; system predictions must also be interpretable to support trust, auditing, and operational decision-making. This paper presents a video-level weapon-related activity classification framework based on a convolutional neural network–long short-term memory (CNN-LSTM) architecture, augmented with explainable artificial intelligence (XAI) techniques. The proposed approach models both spatial appearance and temporal dynamics by extracting frame-level features using a convolutional backbone and learning motion patterns across time using an LSTM network. Model performance is evaluated on a held-out test set using accuracy, precision, recall, and F1-score, achieving a substantial classification performance under controlled experimental conditions, where Grad-CAM++ performed the best in both computability and interpretability. To enhance transparency, gradient-based Class Activation Mapping (CAM) techniques, including Grad-CAM, Grad-CAM++, and Eigen-CAM, are employed to visualize spatial regions contributing to model predictions. Results demonstrate that the proposed framework effectively distinguishes weapon-related scenarios from non-violent activities while providing interpretable visual explanations. The findings highlight the feasibility of integrating explainability into video-based weapon detection pipelines and underscore the importance of transparent AI systems in security-critical applications.