Our work, MAPS, has been accepted to ECCV 2026. The paper reveals fundamental limitations of relying on universal blacklists for safety guidance in generative models.