MIRAGE demonstrates how multimodal language models can be manipulated through narrative immersion techniques, bypassing safety mechanisms designed to prevent harmful outputs.

• Creates jailbreak attacks using environment-role-action triplets to manipulate AI reasoning
• Achieves 84.1% success rate against leading multimodal AI systems including GPT-4V
• Exploits cross-modal reasoning capabilities through carefully crafted visual and textual inputs
• Provides a framework for identifying and addressing critical security vulnerabilities

This research highlights urgent security concerns as multimodal AI systems become more widespread, demonstrating the need for more robust safety measures that protect against sophisticated cross-modal attacks.

MIRAGE: Multimodal Immersive Reasoning and Guided Exploration for Red-Team Jailbreak Attacks