LLM-based phishing schemes are rapidly proliferating, creating a new and highly deceptive class of cyberattacks that traditional detection systems fail to capture. Unlike classical phishing, emails generated by large language models exhibit human-like tone, contextual logic, and linguistic subtlety, allowing them to bypass keyword and rule-based filters. This research presents a novel AI-driven email threat detection architecture that integrates multi-modal transformers, ensemble learning, adversarial robustness, and explainable AI (XAI) to effectively address these advanced and adaptive phishing threats. The proposed system analyzes both textual content and metadata-including email headers, structural patterns, and URL-domain reputations-allowing for deep contextual understanding across multiple modalities. A multi-modal transformer encoder captures semantic and structural cues, while adversarial training improves resilience against LLMgenerated and perturbed phishing samples. Predictions from fine-tuned transformer models are combined with TF-IDF classifiers and URL-risk scoring models using a stacking and soft-voting ensemble strategy, providing reliable performance across diverse attack types. Preliminary results demonstrate that the multi-modal and adversarial training approach increases detection accuracy, improves generalization to unseen phishing variants, and strengthens robustness against adversarial manipulations. Furthermore, an explainability layer using SHAP and LIME provides feature-level insights, and a Streamlit visualization dashboard delivers actionable, interpretable feedback for cybersecurity analysts. The novelty of this research lies in bridging cybersecurity with mathematically grounded explainability, offering a system that is not only accurate and robust but also transparent and auditable. By combining advanced AI reasoning with interpretable mathematical frameworks, this work establishes a new benchmark for trustworthy phishing detection in modern cybersecurity environments.