Large Language Models (LLMs) offer transformative potential for healthcare, but their deployment poses critical privacy risks. Although existing privacy-preserving architectures provide foundational safeguards, they often rely on a single defense mechanism, leaving them vulnerable to multi-stage attacks. To address this gap, we introduce a multi-layered, privacy-enforcing framework that safeguards data while maintaining high utility. Our approach combines proactive prevention, reactive validation, and long-term adaptation. For proactive prevention, we fine-tune a Flan-T5 model with a self-attention masking layer. Finally, for long-term adaptation, an asynchronous Reinforcement Learning with Human Feedback (RLHF) loop iteratively refines the model’s privacy compliance, strengthening its resistance to evolving adversarial attacks. We evaluated our framework against role-based queries and adversarial jailbreak prompts. Our integrated system achieved 90.13% accuracy while reducing privacy leakage to just 3.91%, delivering a significant security improvement over conventional models. Attention heatmap analysis further confirmed that self-attention masking effectively suppresses access to unauthorized tokens. Our findings demonstrate that a multi-layered defense is critical for safe, regulatory-compliant LLM deployment in sensitive domains.