Multimodal Large Language Models (MLLMs) are susceptible to a heuristic-induced multimodal risk distribution jailbreak attack. This sophisticated attack method effectively bypasses existing safety mechanisms by dispersing malicious prompts across both text and image modalities. This distribution strategy ensures that the harmful intent is not detectable when examining either modality in isolation. An auxiliary Large Language Model (LLM) is employed to generate prompts that strategically guide the target MLLM to reconstruct the malicious prompt, ultimately resulting in the production of the intended harmful output. Mitigation steps: **For AI Developers:** * Implement advanced multimodal safety mechanisms to effectively detect and mitigate malicious prompts across various modalities. * Develop sophisticated models capable of detecting harmful intent, ensuring they can identify threats distributed across different modalities. **For Model Trainers/Fine-tuners:** * Enhance methods for identifying and neutralizing prompts that enhance understanding or induce unintended behaviors within the model. * Conduct further research to develop effective countermeasures against vulnerabilities, focusing on areas not covered by existing mitigation strategies.