# Towards Effective MLLM Jailbreaking Through Balanced On-Topicness and OOD-Intensity

arXiv: https://arxiv.org/abs/2508.09218 · HTML: https://mlsanity.com/paper/2508.09218/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Placed in MLLM jailbreaks and adversarial robustness, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 25 of its 30 references that are in the map.
- First posted on arXiv 11 Aug 2025.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM jailbreaks and adversarial robustness](https://mlsanity.com/mllm/mllm-jailbreaks-and-adversarial-robustness/)
- LLM safety and security / [Vision-language and text-to-image safety](https://mlsanity.com/safety/vision-language-and-text-to-image-safety/)

## Builds on (most cited first, 25 of 25)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Qwen2.5-VL Technical Report](https://mlsanity.com/paper/2502.13923/) (2025)
- [Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities](https://mlsanity.com/paper/2507.06261/) (2025)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [PaLM-E: An Embodied Multimodal Language Model](https://mlsanity.com/paper/2303.03378/) (2023)
- [InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models](https://mlsanity.com/paper/2504.10479/) (2025)
- [Jailbreaking Black Box Large Language Models in Twenty Queries](https://mlsanity.com/paper/2310.08419/) (2023)
- [FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space](https://mlsanity.com/paper/2506.15742/) (2025)
- [ShareGPT4V: Improving Large Multi-Modal Models with Better Captions](https://mlsanity.com/paper/2311.12793/) (2023)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [Aligning Large Multimodal Models with Factually Augmented RLHF](https://mlsanity.com/paper/2309.14525/) (2023)
- [Tree of Attacks: Jailbreaking Black-Box LLMs Automatically](https://mlsanity.com/paper/2312.02119/) (2023)
- [NExT-GPT: Any-to-Any Multimodal LLM](https://mlsanity.com/paper/2309.05519/) (2023)
- [DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding](https://mlsanity.com/paper/2412.10302/) (2024)
- [RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback](https://mlsanity.com/paper/2312.00849/) (2023)
- [FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts](https://mlsanity.com/paper/2311.05608/) (2023)
- [Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming](https://mlsanity.com/paper/2501.18837/) (2025)
- [BLIVA: A Simple Multimodal LLM for Better Handling of Text-Rich Visual Questions](https://mlsanity.com/paper/2308.09936/) (2023)
- [Auditing language models for hidden objectives](https://mlsanity.com/paper/2503.10965/) (2025)
- [Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld](https://mlsanity.com/paper/2311.16714/) (2023)
- [Towards End-to-End Embodied Decision Making via Multi-modal Large Language Model: Explorations with GPT4-Vision and Beyond](https://mlsanity.com/paper/2310.02071/) (2023)
- [Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy](https://mlsanity.com/paper/2503.20823/) (2025)
- [Distraction is All You Need for Multimodal Large Language Model Jailbreaking](https://mlsanity.com/paper/2502.10794/) (2025)
- [PiCo: Jailbreaking Multimodal Large Language Models via Pictorial Code Contextualization](https://mlsanity.com/paper/2504.01444/) (2025)

## Built on by (most cited first, 4 of 4)

- [ProEvent: An Event-centric Benchmark for Proactive Agents](https://mlsanity.com/paper/2607.17701/) (2026)
- [MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs](https://mlsanity.com/paper/2603.00565/) (2026)
- [Overloading Large Vision-Language Models for Jailbreaking](https://mlsanity.com/paper/2607.02961/) (2026)
- [Which Image Property Carries the Jailbreak? A Controlled Dissection of Image-to-Text Jailbreaks](https://mlsanity.com/paper/2610.07009/) (2026)

## Abstract (arXiv)

Multimodal large language models (MLLMs) are widely used in vision-language reasoning tasks. However, their vulnerability to adversarial prompts remains a serious concern, as safety mechanisms often fail to prevent the generation of harmful outputs. Although recent jailbreak strategies report high success rates, many responses classified as "successful" are actually benign, vague, or unrelated to the intended malicious goal. This mismatch suggests that current evaluation standards may overestimate the effectiveness of such attacks. To address this issue, we introduce a four-axis evaluation framework that considers input on-topicness, input out-of-distribution (OOD) intensity, output harmfulness, and output refusal rate. This framework identifies truly effective jailbreaks. In a substantial empirical study, we reveal a structural trade-off: highly on-topic prompts are frequently blocked by safety filters, whereas those that are too OOD often evade detection but fail to produce harmful content. However, prompts that balance relevance and novelty are more likely to evade filters and trigger dangerous output. Building on this insight, we develop a recursive rewriting strategy called Balanced Structural Decomposition (BSD). The approach restructures malicious prompts into semantically aligned sub-tasks, while introducing subtle OOD signals and visual cues that make the inputs harder to detect. BSD was tested across 13 commercial and open-source MLLMs, where it consistently led to higher attack success rates, more harmful outputs, and fewer refusals. Compared to previous methods, it improves success rates by 67% and harmfulness by 21%, revealing a previously underappreciated weakness in current multimodal safety systems.