Multimodal Large Language Models (MLLMs) are susceptible to a sophisticated attack vector that exploits narrative-driven visual storytelling and role immersion to bypass inherent safety protocols. This attack, known as MIRAGE, strategically decomposes malicious queries into triplets consisting of environment, character, and activity. It then generates a sequence of images and text prompts to guide the MLLM through a misleading narrative, ultimately provoking harmful outputs. The attack effectively leverages the MLLM's cross-modal reasoning capabilities and its vulnerability to persona-based manipulation. Mitigation steps: **For AI Developers:** * Implement pre-screening mechanisms using vision-language models to analyze visual inputs for potentially harmful content before processing by the MLLM. * Develop sophisticated detection methods to identify attempts at role immersion and deceptive storytelling. **For Model Trainers/Fine-tuners:** * Improve the robustness of MLLM safety mechanisms to handle multi-turn interactions and narrative contexts. * Enhance the training data used for MLLM safety reinforcement by including examples of narrative-driven attacks.