Back to Paper
cs.CRcs.AIcs.MM
Local ID: 2603.21697v2
AI Summary: gemma4:e4b
Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models
By Rui Yang Tan, Yujia Hu, Roy Ka-Wei Lee
Revision History Timeline
v13/23/2026
3/23/2026
“31 pages”
v24/23/2026
4/23/2026
“Code released at: https://github.com/Social-AI-Studio/ComicJailbreak”
★ Version indexed in ExplorerComparing v1 vs v2
Green = Added • Red = Removed
Title Comparison
Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models
Authors Comparison
No author changes.
v1 Comment
“31 pages”
v2 Comment
“Code released at: https://github.com/Social-AI-Studio/ComicJailbreak”
Abstract Word Diff
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, we show that current safety evaluators can be unreliable on sensitive but non-harmful content. Our findings highlight the need for safety alignment robust to narrative-driven multimodal jailbreaks.