The extensive adoption of SMS for communication has resulted in an increase in phishing and misleading messages that jeopardize user security. This study introduces a classification system to sort SMS messages into three risk categories: safe, potentially harmful, and clearly harmful. The dataset, made up of 10,000 user-submitted SMS messages gathered through an online form, was labeled and analyzed manually. A refined transformer model (LLaMA 3.2) was used for multi-class classification. Performance evaluation included precision, recall, F1-score, and ROC-AUC, resulting in an overall accuracy of 81 %. Tools like word clouds and confusion matrices were utilized for result interpretation. The results show that transformer models can accurately identify SMS risk categories. The model demonstrates potential, but it has limitations, such as missing contextual and behavioral metadata. Future work will incorporate explainable AI (XAI) methods to improve clarity and foster user confidence.