[2603.21697] Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models

Computer Science > Cryptography and Security arXiv:2603.21697 (cs) [Submitted on 23 Mar 2026] Title:Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models Authors:Rui Yang Tan, Yujia Hu, Roy Ka-Wei Lee View a PDF of the paper titled Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models, by Rui Yang Tan and 2 other authors View PDF HTML (experimental) Abstract:Multimodal Large Language Models (MLLMs) extend text-only LLMs with visual reasoning, but also introduce new safety failure modes under visually grounded instructions. We study comic-template jailbreaks that embed harmful goals inside simple three-panel visual narratives and prompt the model to role-play and "complete the comic." Building on JailbreakBench and JailbreakV, we introduce ComicJailbreak, a comic-based jailbreak benchmark with 1,167 attack instances spanning 10 harm categories and 5 task setups. Across 15 state-of-the-art MLLMs (six commercial and nine open-source), comic-based attacks achieve success rates comparable to strong rule-based jailbreaks and substantially outperform plain-text and random-image baselines, with ensemble success rates exceeding 90% on several commercial models. Then, with the existing defense methodologies, we show that these methods are effective against the harmful comics, they will induce a high refusal rate when prompted with benign prompts. Finally, using automatic judging and targeted human evaluation...