Large-scale black-box testing reveals that inexpensive, gradient-free image transformations can bypass three major commercial multimodal moderation APIs. These attacks remain effective across various harm categories and transformation intensities without requiring specialized knowledge.
HOW THIS AFFECTS YOU
●
builderYou should not rely solely on commercial moderation APIs for safety, as they are vulnerable to simple visual edits.
●
policyCurrent multimodal safety standards are insufficient to prevent bypasses via basic image manipulation.