We propose adversarial paraphrasing: a training-free, transferable attack that universally humanizes AI-generated text by using an off-the-shelf LLM as a paraphraser, guided by an AI text detector.