Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text