📝 SummarXiv

Abstract

Large Language Models (LLMs) watermarking has shown promise in detecting AI-generated content and mitigating misuse, with prior work claiming robustness against paraphrasing and text editing. In this paper, we argue that existing evaluations are not sufficiently adversarial, obscuring critical vulnerabilities and overstating the security. To address this, we introduce adaptive robustness radius, a formal metric that quantifies watermark resilience against adaptive adversaries. We theoretically prove that optimizing the attack context and model parameters can substantially reduce this radius, making watermarks highly susceptible to paraphrase attacks. Leveraging this insight, we propose RLCracker, a reinforcement learning (RL)-based adaptive attack that erases watermarks while preserving semantic fidelity. RLCracker requires only limited watermarked examples and zero access to the detector. Despite weak supervision, it empowers a 3B model to achieve 98.5% removal success and an average 0.92 P-SP score on 1,500-token Unigram-marked texts after training on only 100 short samples. This performance dramatically exceeds 6.75% by GPT-4o and generalizes across five model sizes over ten watermarking schemes. Our results confirm that adaptive attacks are broadly effective and pose a fundamental threat to current watermarking defenses.