Phishing attacks continue to pose a serious cybersecurity threat by exploiting social engineering techniques to deceive users into disclosing sensitive information. These attacks commonly rely on malicious Uniform Resource Locators (URLs), often hosted on newly registered or parked domains to evade traditional blacklist-based detection systems. Early identification of such URLs is essential to reduce financial losses and identity theft. This paper presents a machine learning–based framework for the early detection of malicious URLs, with particular emphasis on newly registered and parked domains. A dataset comprising 211,659 URLs was constructed using real-time SSL certificate monitoring, popular domain listings, and verified phishing reports. The proposed approach incorporates data preprocessing, URL-based feature extraction, class balancing, and model optimization. Experimental results demonstrate that the Light Gradient Boosting Machine (LGBM) classifier achieves a recall of 96.02% and an accuracy of 97.28% using 10-fold cross-validation. Feature selection techniques further reduce model complexity while maintaining detection performance, enabling practical deployment. The framework provides a proactive and scalable solution for phishing prevention and brand protection in sectors such as banking and e-commerce.