Cyberbullying on social media poses significant psychological risks, yet most detection systems oversimplify the task by focusing on binary classification, ignoring nuanced categories like passive-aggressive remarks or indirect slurs. To address this gap, we propose a hybrid framework combining Term Frequency-Inverse Document Frequency (TF-IDF), word-to-vector (Word2Vec), and Bidirectional Encoder Representations from Transformers (BERT) based models for multi-class cyberbullying detection. Our approach integrates TF-IDF for lexical specificity and Word2Vec for semantic relationships, fused with BERT’s contextual embeddings to capture syntactic and semantic complexities. We evaluate the framework on a publicly available dataset of 47,000 annotated social media posts across five cyberbullying categories: age, ethnicity, gender, religion, and indirect aggression. Among BERT variants tested, BERT Base Un-Cased achieved the highest performance with 93% accuracy (standard deviation ±1% across 5-fold cross-validation) and an average AUC of 0.96, outperforming standalone TF-IDF (78%) and Word2Vec (82%) models. Notably, it achieved near-perfect AUC scores (0.99) for age and ethnicity-based bullying. A comparative analysis with state-of-the-art benchmarks, including Generative Pre-trained Transformer 2 (GPT-2) and Text-to-Text Transfer Transformer (T5) models highlights BERT’s superiority in handling ambiguous language. This work advances cyberbullying detection by demonstrating how hybrid feature extraction and transformer models improve multi-class classification, offering a scalable solution for moderating nuanced harmful content.