RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks

Hanbo Huang, Yiran Zhang, Hao Zheng, Xuan Gong, Yihan Li, Lin Liu, Zhuotao Liu, Shiyu Liang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:48441-48483, 2026.

Abstract

Large language model (LLM) 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 the adaptive robustness radius, a formal metric that quantifies the worst-case resilience of watermarks against adaptive adversaries. By lifting the paraphrase space into a KL-divergence ball, we approximate this radius and theoretically demonstrate that optimizing the attack context and model parameters can significantly reduce the approximated radius, making watermarks highly vulnerable 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 with minimal semantic shift 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 code is available in this repository.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-huang26ch, title = {{RLC}racker: Evaluating the Worst-Case Vulnerability of {LLM} Watermarks with Adaptive {RL} Attacks}, author = {Huang, Hanbo and Zhang, Yiran and Zheng, Hao and Gong, Xuan and Li, Yihan and Liu, Lin and Liu, Zhuotao and Liang, Shiyu}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {48441--48483}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/huang26ch/huang26ch.pdf}, url = {https://proceedings.mlr.press/v306/huang26ch.html}, abstract = {Large language model (LLM) 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 the adaptive robustness radius, a formal metric that quantifies the worst-case resilience of watermarks against adaptive adversaries. By lifting the paraphrase space into a KL-divergence ball, we approximate this radius and theoretically demonstrate that optimizing the attack context and model parameters can significantly reduce the approximated radius, making watermarks highly vulnerable 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 with minimal semantic shift 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 code is available in this repository.} }
Endnote
%0 Conference Paper %T RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks %A Hanbo Huang %A Yiran Zhang %A Hao Zheng %A Xuan Gong %A Yihan Li %A Lin Liu %A Zhuotao Liu %A Shiyu Liang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-huang26ch %I PMLR %P 48441--48483 %U https://proceedings.mlr.press/v306/huang26ch.html %V 306 %X Large language model (LLM) 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 the adaptive robustness radius, a formal metric that quantifies the worst-case resilience of watermarks against adaptive adversaries. By lifting the paraphrase space into a KL-divergence ball, we approximate this radius and theoretically demonstrate that optimizing the attack context and model parameters can significantly reduce the approximated radius, making watermarks highly vulnerable 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 with minimal semantic shift 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 code is available in this repository.
APA
Huang, H., Zhang, Y., Zheng, H., Gong, X., Li, Y., Liu, L., Liu, Z. & Liang, S.. (2026). RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:48441-48483 Available from https://proceedings.mlr.press/v306/huang26ch.html.

Related Material