Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models

TL;DR

This paper demonstrates that multimodal large language models are vulnerable to visual narrative-based jailbreaks, exposing safety risks and highlighting the need for more robust safety alignment methods.

Contribution

Introduces ComicJailbreak, a new benchmark for visual narrative jailbreaks, and evaluates the effectiveness of current safety defenses against these multimodal attacks.

Findings

ComicJailbreak contains 1,167 attack instances across 10 harm categories.

State-of-the-art models show over 90% success rate in visual jailbreak attacks.

Current safety evaluators are unreliable on sensitive but non-harmful content.

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, 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.