Web vulnerability detection has increased recently comparing to years before, moving from manual inspection and signature based methods to advanced machine learning techniques. Traditional approaches, such as manual code reviews, were effective but it used to take more time, error prone, and difficult to scale. As web applications became more complex, the need for automated, efficient solutions has grown. This paper examines the shift from static analysis and expert driven detection to AI powered methods. In the past, machine learning models in this field started with basic statistical techniques and years later, it reached to deep learning, like convolutional neural networks (CNNs) and recurrent neural networks (RNNs). These models have made it possible of detection accuracy, decreasing false negatives, and enabling large number of vulnerability identification. Systems like VulDeePecker and \(\mu \) VulDeePecker are milestones in this field, utilizing bidirectional LSTMs (BLSTMs) and hybrid models to detect vulnerabilities without having to manually engineer features. Our paper deep dive in the current state of ML in the field of vulnerability detection, also it shown the challenges of involving AI, additionally future directions to make systems more secure and safe. ML is doing a revolution of web security by building scalable, automated and adaptive remediation to include concerned cyber-threats.