The rapid growth of toxic content on social media is a critical issue, but existing detection models often fail because they ignore the message's context, leading to poor accuracy, especially with veiled aggression. This work proposes a Multi-dimensional Context-aware Model (MCDM) to solve this. The MCDM integrates the message text with its discursive, user, and temporal context using a modified Transformer architecture (e.g., BERT/RoBERTa). This novel approach is theoretically justified to significantly improve the identification of toxic content, including sarcastic and subtle forms of aggression.