Cyberbullying is a growing concern across social media platforms, necessitating advanced detection mechanisms to mitigate its impact. Traditional machine learning models often struggle with understanding contextual dependencies and ensuring model interpretability. This paper proposes a hybrid deep learning approach that combines BERT and RoBERTA for feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) networks for sequential dependency learning. To enhance interpretability, attention mechanisms such as Self-Attention and Bahdanau Attention are integrated, allowing the model to focus on crucial words contributing to classification. The proposed system aims to improve accuracy, scalability, and explainability while addressing key challenges in cyberbullying detection. This research lays the groundwork for developing more transparent and effective AI-driven moderation systems for online safety.