Abstract
Researchers have identified a significant vulnerability in Multimodal Large Language Models (MLLMs) termed Sequential Comic Jailbreak (SCJ). This attack exploits the narrative reasoning capabilities of models by decomposing malicious instructions into a series of individually innocuous comic panels. By leveraging the model's drive for narrative coherence, SCJ achieves an average attack success rate of 83.5%, outperforming previous single-frame visual jailbreak methods by approximately 46%.
Introduction
As Multimodal Large Language Models (MLLMs) like GPT-4o and Gemini become increasingly integrated into digital ecosystems, ensuring their safety alignment is critical. While safety measures are robust in the text domain, visual jailbreaks—particularly those involving sequential reasoning—remain an underexplored threat. This research investigates whether an MLLM’s focus on "connecting the dots" in a visual narrative can be used to bypass safety filters, a phenomenon the authors call asymmetric alignment.
Methodology: The SCJ Framework
The authors propose a four-phase framework designed to transform a harmful query into a structured visual story that evades detection:
- Query Intention Extraction: An auxiliary LLM decomposes a harmful prompt into core intent, roles, resources, and implementation steps.
- Story Script Creation: These components are woven into a comic script where each individual scene appears benign.
- Comics Generation: A diffusion model renders the script. Because individual panels are innocuous, they bypass the diffusion model's internal safety filters.
- Target Model Attack: The sequence is presented to the target MLLM with prompts encouraging narrative reconstruction or detective-style analysis.
Experimental Setup
The study evaluated SCJ using 1,556 queries across 11 safety-critical categories from two major benchmarks:
- MM-SafetyBench: 806 queries covering Illegal Activity, Hate Speech, Physical Harm, Fraud, Sexual Content, and Privacy Violation.
- HADES: 750 queries covering Violence, Financial Crime, Privacy, Self-Harm, and Animal Abuse.
Results and Analysis
The SCJ attack demonstrated a massive success rate, achieving an average 83.5% Attack Success Rate (ASR) across 11 state-of-the-art models. Open-source models like Qwen3-VL and DeepSeek-VL2 were particularly vulnerable, often exceeding 95% ASR. Commercial models also showed high susceptibility, with GPT-4o and Gemini 2.5 Pro ranging between 78% and 84%.
Key insights from the results include:
- Narrative Over Scrutiny: MLLMs prioritize story completion. When forced to infer implicit information between panels, they often fail to apply safety scrutiny to the emergent harmful intent.
- Capability vs. Safety: More advanced models, which possess superior narrative reasoning skills, can be more susceptible to SCJ as they are better at reconstructing the "hidden" harmful message.
- Category Sensitivity: Action-oriented categories like "Illegal Activity" or "Fraud" are most vulnerable due to their natural mapping to step-by-step comic sequences.
Conclusion and Future Work
The vulnerability exposed by SCJ highlights a critical gap in current MLLM defenses, which are often "frame-blind" and analyze images in isolation. To counter these threats, the authors suggest developing narrative-aware safety guardrails that analyze temporal patterns across multiple inputs. As AI moves toward more complex video-language models, accounting for structured storytelling in safety evaluations will be essential.
Summary updated