Challenge: Existing jailbreak methods only use a single image, restricting the attack space . Existing frameworks only use single image to distribute harmful requests across multiple images .
Approach: They propose a compositional jailbreak framework that leverages Distributed instruction, Multimodal evidence and a Number chain task to fully enhance the jailbreak performance.
Outcome: The proposed framework achieves attack success rates of over 90% on GPT-4o, Gemini-2.5-pro and Claude Sonnet 4 .

Similar Papers

From LLMs to MLLMs: Exploring the Landscape of Multimodal Jailbreaking (2024.emnlp-main)

Copied to clipboard

Challenge: Recent advances in Large Language Models (LLMs) have demonstrated remarkable performance across various tasks, effectively following instructions to meet diverse user needs.
Approach: They propose a framework for evaluation benchmarks and attack techniques for LLMs and MLLMs to enhance their security.
Outcome: The proposed frameworks have been exploited to exploit the weaknesses of LLMs and MLLMs.
MLLM-Protector: Ensuring MLLM’s Safety without Hurting Performance (2024.emnlp-main)

Copied to clipboard

Challenge: MLLMs are deployed on limited image-text pairs, which makes them more vulnerable to catastrophic forgetting of their original abilities during safety fine-tuning.
Approach: They propose a plug-and-play strategy that detects harmful visual inputs and transforms harmful ones into harmless ones.
Outcome: The proposed approach mitigates the risks posed by malicious visual inputs without compromising the original performance of MLLMs.
Cross-modality Information Check for Detecting Jailbreaking in Multimodal Large Language Models (2024.findings-emnlp)

Copied to clipboard

Challenge: Multimodal Large Language Models (MLLMs) are susceptible to jailbreak attacks, authors say . multimodal information increases the risk of attacks, but also provides additional data .
Approach: They propose a jailbreaking detector that detects maliciously perturbed image inputs . cross-modality information detector is designed to detect cross-modal similarity between harmful queries and adversarial images.
Outcome: a new tool can detect maliciously perturbed image inputs without modification or computation cost.
Jailbreaking Multimodal Large Language Models using Multi-Clip Video (2026.acl-long)

Copied to clipboard

Challenge: Existing studies show that video inputs can bypass safety alignment, yet it remains unclear which properties of video input induce this vulnerability.
Approach: They propose a simple image-based defense that mitigates the vulnerability of MLLMs by analyzing video inputs.
Outcome: The proposed defense leverages the relative robustness of the image modality.
Reference Attack: A New Cross-Modal Jailbreaking Attack against Multimodal Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) have raised significant safety concerns about generated content, drawing attention from both academia and industry.
Approach: They propose a reference-guided cross-modal jailbreak method that enhances existing prompt-to-image injection attacks by exploiting MLLMs’ semantic reconstruction capabilities.
Outcome: The proposed method achieves an attack success rate of over 93% on leading MLLMs including ChatGPT, Gemini, Claude, and the widely used open-source LLaMA model.
SafeSteer: A Decoding-level Defense Mechanism for Multimodal Large Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Existing defense methods rely on fine-tuning or inefficient post-hoc interventions, limiting their ability to address novel attacks.
Approach: They propose a decoding-level defense mechanism that employs a lightweight discriminator to iteratively steer the decoding process toward safety.
Outcome: The proposed method improves safety performance by up to 33.40% without fine-tuning on multiple MLLMs.
GAMBIT: A Gamified Jailbreak Framework for Multimodal Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing attacks focus on increasing the complexity of the modified visual task and do not explicitly leverage the model’s own reasoning incentives.
Approach: They propose a framework that decomposes and reassembles harmful visual semantics and constructs a gamified scene that drives the model to explore, reconstruct intent and answer as part of winning the game.
Outcome: Experiments on reasoning and non-reasoning MLLMs show that the proposed framework outperforms baseline models on both vision and text.
FC-Attack: Jailbreaking Multimodal Large Language Models via Auto-Generated Flowcharts (2025.findings-emnlp)

Copied to clipboard

Challenge: Recent research shows that multimodal large language models are vulnerable to jailbreak attacks .
Approach: They propose a jailbreak attack method based on auto-generated flowcharts . the flowchartings are then combined with a benign textual prompt to execute the attack .
Outcome: The proposed method achieves an attack success rate of up to 96% via images and 78% via videos across multiple MLLMs.
Jailbreak Large Vision-Language Models Through Multi-Modal Linkage (2025.acl-long)

Copied to clipboard

Challenge: Existing methods to jailbreak large vision-language models fail against cutting-edge models such as GPT-4o, despite having undergone safety alignment training.
Approach: They propose a new framework for jailbreaking large vision-language models that uses an encryption-decryption process to mitigate the over-exposure of harmful information.
Outcome: The proposed framework jailbreaks GPT-4o with 99.40% success rates on SafeBench, 98.81% on MM-SafeBench and 99.07% on HADES-Dataset.
Exploring Compositional Generalization of Multimodal LLMs for Medical Imaging (2025.acl-long)

Copied to clipboard

Challenge: Current research suggests that multitask training outperforms single-task as different tasks can benefit each other, but they often overlook the internal relationships within these tasks.
Approach: They employ compositional generalization (CG) to examine the generalization of multimodal large language models in medical imaging.
Outcome: The proposed model can understand unseen medical images and is able to perform CG across classification and detection tasks.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations