The rapid evolution of Internet technology has revolutionized information exchange through web applications, yet it has also introduced significant security risks, with cross-site scripting (XSS) attacks ranking as a top threat according to the Open Web Application Security Project (OWASP). XSS attacks can cause session hijacking, data theft, and service disruption, making their detection critical. Traditional machine learning models struggle with the dynamic and evolving nature of cross-site scripts, necessitating more sophisticated detection mechanisms. This paper addresses this challenge by proposing a hybrid deep learning model that combines convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) networks, augmented by self-attention mechanisms and character-level FastText embeddings. This hybrid model leverages CNN ability to capture spatial features and BiLSTM’s strength in handling sequential dependencies in web traffic. Through extensive experimentation, including ablation studies and comparisons with traditional machine learning models, the proposed model demonstrates superior performance, achieving an accuracy of 99.61%, a recall of 99.37%, and an F1 score of 99.61%. Furthermore, the model incorporates online learning techniques to address concept drift and ensure adaptability over time. The contributions of this paper include a novel hybrid architecture for XSS detection, detailed analysis of its performance, and strategies to maintain model effectiveness in dynamic environments.