The recent proliferation of digital communication has created new avenues for social interaction, but it has also amplified exposure to harmful online behavior, particularly in the form of cyberbullying. This type of abuse is increasingly spreading on social media and has been linked to adverse psychological effects on targeted individuals. Although artificial intelligence (AI) and natural language processing (NLP) provide powerful tools for detecting and combating this content, their practical effectiveness is often limited by the complexities of selecting the appropriate models and fine-tuning their hyperparameters. In response to these challenges, this study introduces an innovative AI-based framework for identifying cyberbullying on social media. The approach integrates Word2Vec embeddings to capture the semantic meaning of text and employs the XGBoost algorithm for classification. To optimize performance, a customized version of the reptile search algorithm is applied for hyperparameter tuning. The proposed system is benchmarked against other metaheuristic techniques using a collection of traditional classification metrics. Results from the experiments highlight the robustness of the approach, with the leading model achieving an accuracy of around 93.7%. These outcomes illustrate the promising synergy between NLP techniques and metaheuristic optimization for the effective detection of cyberbullying within digital landscapes.