<p>Accurately classifying password origins is significant for cybersecurity applications such as password strength estimation, attack detection, and forensic investigations. The present work proposes a hybrid ResCNN-BiGRU-Gated Attentive Pooling (GAP) model that leverages character-level embeddings to detect both local and sequential patterns in passwords. The dataset includes a balanced mix of real and synthetically generated passwords, encompassing human-generated, rule-based, Markov, and keyboard walk-based types. The proposed framework integrates Residual-CNN blocks for extracting multi-scale local features, BiGRU layers for modeling long-range sequential relationships, and GAP for emphasizing discriminative features while reducing noise. Hyperparameters were optimized using the Augmented Weighted K-means Grey Wolf Optimizer (AWK-GWO), ensuring maximized accuracy and generalization. Comparative experiments with baseline models demonstrate the effectiveness of the proposed approach, achieving a testing accuracy of 0.9521, F1-macro of 0.952, ROC-AUC of 0.9959, MCC of 0.9401, and Log-Loss of 0.1329. Confusion matrices, ROC curves, and training plots indicate consistent class separation, stable convergence, and well-calibrated probabilities. The combination of convolution, recurrence, and attention mechanisms enables the model to capture intricate patterns in password generation more effectively than individual architectures. This framework offers a robust tool for identifying password origins, supporting cybersecurity forensics, attack mitigation, and security surveillance.</p>