In the era of digital transformation, alongside the rapid devel-opment of the Internet and online applications, phishing at-tacks targeting users through malicious URLs have been in-creasingly prevalent. Traditional methods for detecting mali-cious URLs often rely on blacklist-based techniques. Howev-er, these techniques have significant limitations as they cannot identify new URLs. Many machine learning-based approaches have been researched and implemented to overcome these shortcomings. This paper proposes an improved two-stage deep learning model using an Autoencoder network for phish-ing URL detection. The proposed approach will be evaluated and tested on the standard UCI's URL Phishing dataset, achieving better results in most measure metrics.