Researchers describe the Jailbreak Connectivity method for bypassing the safety guardrails of multimodal AI models
The new Jailbreak Connectivity method achieves, according to the authors, an 84–98% success rate at bypassing the safety guardrails of multimodal language models on the MM-SafetyBench benchmark, including black-box tests.
A research team presented the Jailbreak Connectivity (JC) method, a framework for attacking the safety guardrails of multimodal language models (MLLM) using modified image inputs. Unlike previous approaches that optimize isolated adversarial examples, JC is based on the observation that effective adversarial inputs can be connected through low-loss regions in image space. The method combines path sampling for diversity of generated jailbreak images, auxiliary classifiers for transferring the attack between models, and a distribution-level objective function for universal attacks across different harmful queries.
According to the authors, on the MM-SafetyBench benchmark JC achieves an average attack success rate of 84.01%, 85.77%, and 97.95% on the Qwen3-VL-8B-Instruct, InternVL3.5-8B, and Gemma-3-12B-IT models, respectively, thereby outperforming the strongest baseline methods by up to 68.36 percentage points. In black-box tests (without knowledge of the model's internal structure), the method, according to the authors, improves attack success rate over the FigStep method by 24.69 and 27.34 percentage points on the models referred to in the article as GPT-5.6 and Gemini 3.1 Pro.
The authors state that the results reveal structured vulnerabilities in the safety alignment of multimodal models and are intended to serve more thorough security evaluation. For details on the methodology and other related sources, see the source article.
Why it matters
The result shows that the safety guardrails of a number of multimodal models can be bypassed using modified images with a documented high success rate, even without access to the model's internal structure. For companies deploying multimodal models in their products, this represents a concrete risk of misuse and the need to incorporate similar attacks into security testing before deployment.
Relevant practical impact
What this means
For a business
Companies deploying multimodal language models (e.g., for image analysis in products) must account for the fact that existing safety guardrails can be bypassed by attacks via image input with a high success rate, even in black-box models accessible only through an API.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.