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

2025-03-24

AI Summary

Multimodal large language models (MLLMs) exhibit critical security vulnerabilities in cross-modal reasoning, rendering them susceptible to multimodal jailbreaking attacks that circumvent existing safety mechanisms and elicit harmful outputs. Method: We propose the first narrative-driven, role-immersive multimodal jailbreaking framework. It constructs interleaved image-text, multi-turn visual story sequences via environment-role-action triples to progressively weaken model defenses. By innovatively integrating role immersion with structured semantic reconstruction, our method reveals that narrative cues can spontaneously activate inherent model biases—overcoming limitations of text-only jailbreaking. Technically, it unifies Stable Diffusion-based image generation, multi-turn visual narrative modeling, cross-modal contextual guidance, and decomposition of toxic queries into semantic triples. Contribution/Results: Our framework achieves state-of-the-art attack success rates across six mainstream MLLMs and multiple benchmarks, outperforming the best prior baseline by 17.5%, thereby exposing systemic gaps in current cross-modal safety mechanisms.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

While safety mechanisms have significantly progressed in filtering harmful text inputs, MLLMs remain vulnerable to multimodal jailbreaks that exploit their cross-modal reasoning capabilities. We present MIRAGE, a novel multimodal jailbreak framework that exploits narrative-driven context and role immersion to circumvent safety mechanisms in Multimodal Large Language Models (MLLMs). By systematically decomposing the toxic query into environment, role, and action triplets, MIRAGE constructs a multi-turn visual storytelling sequence of images and text using Stable Diffusion, guiding the target model through an engaging detective narrative. This process progressively lowers the model's defences and subtly guides its reasoning through structured contextual cues, ultimately eliciting harmful responses. In extensive experiments on the selected datasets with six mainstream MLLMs, MIRAGE achieves state-of-the-art performance, improving attack success rates by up to 17.5% over the best baselines. Moreover, we demonstrate that role immersion and structured semantic reconstruction can activate inherent model biases, facilitating the model's spontaneous violation of ethical safeguards. These results highlight critical weaknesses in current multimodal safety mechanisms and underscore the urgent need for more robust defences against cross-modal threats.

Problem

• Exploits multimodal reasoning to bypass MLLM safety mechanisms
• Uses narrative-driven context to lower model defenses
• Activates model biases to violate ethical safeguards

Innovation

• Multimodal jailbreak using narrative-driven context
• Stable Diffusion for visual storytelling sequences
• Role immersion to bypass safety mechanisms