A new preprint on arXiv describes a jailbreak attack that targets large image editing models by distributing harmful intent across both textual and visual inputs. The authors argue that these models increasingly infer editing intent from multiple modalities, creating a safety risk that has not been fully explored.
The attack, which the paper calls a cross-modal jailbreak, splits a harmful request into parts that individually may appear benign, but together form a malicious instruction when processed by the model. Because the model combines information from text and image, the harmful intent is only apparent after fusion, potentially bypassing safety filters that check each modality separately.
The authors formalize this attack scenario, though the abstract does not provide experimental details or mitigation strategies. The work highlights the need for safety mechanisms that consider the joint meaning of multimodal inputs, rather than evaluating text and image in isolation.