<p>Cross-platform generalization results.</p>

The explosive rise of short-form video platforms such as Instagram Reels, TikTok, and YouTube Shorts has transformed digital expression while intensifying the spread of cyberbullying. Unlike video abuse conveys multimodal cues visual, text-based harassment, auditory, and textual that challenge conventional detection methods. This study presents a hybrid deep-learning framework that integrates Convolutional Neural Networks (CNNs) for spatial features, Bidirectional Long Short Term Memory (BiLSTM) networks for temporal acoustic patterns, and a Transformer-based textual encoder to analyze synchronized video, audio, and caption streams. A semantic-consistency validation layer enforces cross-modal alignment using attention-based similarity constraints, ensuring that incongruent cues are penalized during classification. Experiments on two benchmark datasets, CAVD and SocialVidMix, demonstrate state-of-the-art performance accuracy 91.6%, precision 89.7%, recall 93.0%, and F1-score 91.3% with consistent results across Instagram, TikTok, and YouTube Shorts. The framework’s, interpretability, robustnessand scalability indicate strong potential for real-time deployment in automated content-moderation systems.