In recent days, phishing on websites has been a major cybersecurity threat. Hence correct detection is very important in real world systems. Traditional binary phishing detection techniques have high false positive and false negative rates in ambiguous cases. In this paper, we propose an explainable tri-class phishing detection framework with quantum inspired QUBO feature selection. Our framework incorporates a “suspicious” class in addition to “phishing” and “legitimate”. This reduces the possibility of selecting wrong decisions in ambiguous circumstances and also optimized the 54 features to 5 which maximizes the relevance and minimizes the redundancy. Gradient Boosting achieves 99.70 % accuracy with a 99.70 % F1-score. In comparison to binary baselines, the proposed tri-class model achieved zero high-confidence false positives and false negative. Robust generalization (σ<0.0002, train-test gap <2.3%) is confirmed by cross-validation.