The rapid digitization of education, finance, and healthcare has substantially broadened the attack surface available to cybercriminals, with malicious URLs acting as key channels for phishing, malware, and spam delivery. This danger is especially pronounced in developing countries such as Ethiopia, where the pace of digital uptake exceeds the development of security frameworks. This study responds to the urgent need for robust, scalable systems for detecting malicious URLs by proposing a deep learning framework that jointly leverages DistilBERT for semantic representation, Convolutional Neural Networks for identifying local patterns, and Bidirectional Long Short-Term Memory networks augmented with attention mechanisms to capture sequential dependencies within URLs. To tackle pronounced class imbalance in the original internal samples, the Synthetic Minority Over-Sampling Technique was applied only to the training subset, while the validation and test sets were kept naturally imbalanced to avoid data leakage. Four required safeguards were implemented to rigorously reduce synthetic pattern over fitting: augmentation restricted to the training set, independent external benchmarking on 50,000 URLs with no overlap, close generalization-gap analysis, and verification of per-class performance consistency. Across an extensive experimental corpus of 499,960 total instances, the proposed architecture attains 99.85% stratified 5-fold cross-validation accuracy and 98.80% external accuracy on real-world test supports from URLHaus and PhishTank, supported by strong macro-average F1-scores of 98.5% and 98.1% respectively. These findings substantiate exceptional, unbiased multi-class discrimination among Safe, Phishing, Malware, Spam, and Other Threat categories, providing a scalable, reliable detection framework for resource-limited cyber security settings.