From Hallucination Detection to Adversarial Defense: A Unified Framework for LVLM and LLM Safety