The JMLLM attack represents a sophisticated hybrid multimodal jailbreaking strategy targeting vulnerabilities in 13 widely-used large language models (LLMs) across text, image, and speech modalities. This attack employs techniques such as alternating translation, word encryption, feature collapse in images, and harmful text injection to circumvent established safety protocols and provoke harmful outputs. The effectiveness of the attack varies among different LLMs and modalities, with certain models demonstrating a heightened susceptibility to these vulnerabilities. Mitigation steps: **For AI Developers:** * Implement a "Harmful Separator" defense mechanism to separate instructions from examples in prompts. * Conduct independent analysis of examples for potential malicious content. **For Model Trainers/Fine-tuners:** * Investigate and develop additional mitigation strategies to enhance protection against malicious content. * Engage in ongoing research to improve defense mechanisms and address vulnerabilities.