Phishing sites are a serious cyber threat as they trick users into revealing sensitive personal data. Conventional detection techniques, including rule-based systems and blocklists, cannot cope with changing phishing tactics. In this paper, a new approach to phishing detection is introduced using a Bidirectional Long Short-Term Memory (BiLSTM) network with an attention mechanism. The suggested model learns and examines URL-based features and identifies forward and backward relationships in data, enhancing classification accuracy. A 30,000 URL-tagged dataset is utilized to train the model, which is then optimized with the help of methods like sequence tokenization, embedding layers, dropout regularization, and class weight balancing to counter data imbalance issues. The BiLSTM layer processes sequential information about URLs in a bidirectional manner, whereas the attention mechanism applies weights to important features differently to ensure the model pays attention to the most critical elements of phishing URLs. The model was tested based on standard performance metrics and has attained an astounding accuracy of 99.22%, precision of 99.1%, recall of 99.3%, and an F1-score of 99.2%, surpassing the traditional approach like Logistic Regression. The model indicates good generalization ability and is possible to be applied in real-time in web security systems. In the future, the use of dynamic data analysis and large datasets will be applied to improve further the detection efficiency and responsiveness against the new emerging phishing attacks.