acceptodds
Under review as a conference paper at ICLR 2027

Randomness Reuse in LLM Watermarks: Security Bounds and Black-Box Spoofing

Abstract

Generative watermarking is increasingly used to identify LLM-generated content. While theoretical security notions and practical attacks have been studied, their relationship and whether attacks exploit general design properties or implementation details remain unclear. We study these questions through formal analysis and black-box spoofing attacks on SynthID-Text. We first connect two watermark security goals usually analyzed separately: undetectability and spoofing resistance, showing that a successful spoofing attack also yields a distinguisher between watermarked and clean LLM generation when one terminal verification query on a fresh candidate is allowed. We then derive a general security bound for context-dependent watermarking schemes in which randomness reuse explicitly degrades security. Our experiments test whether this reuse is exploitable in practice. On SynthID-Text, we developed black-box spoofing attacks that recover watermark bias from model outputs. An implementation-specific weakness in the public Hugging Face implementation further strengthens the attack, yielding joint detection-and-automated-quality rates of \(82.6%\) and \(88.7%\) on Gemma-2B and Gemma-9B prompts, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM watermarking, watermark security, spoofing attacks, SynthID-Text, undetectability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SFfnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how randomness reuse affects the security of context-dependent LLM watermarks. It relates spoofing resistance to undetectability when one terminal verification query is allowed, then derives a security bound involving the probability of reusing a watermark-active context. For SynthID-Text, the authors develop an exact-context attack and a bucket-pooled attack that exploits publicly identifiable hash equivalences in the implementation. On 512 held-out prompts, the pooled attack achieves joint detector acceptance and automated quality pass rates of 82.6% on Gemma-2B and 88.7% on Gemma-9B. The paper also tests the role of these equivalences through controls and a hash patch.

Critical strengths
  • The bucket-pooled attack identifies a concrete implementation weakness and explains how observations can be shared across different contexts without recovering the watermark key.

  • The controls support the proposed mechanism: permuting bucket assignments removes detector acceptance, while patching the hash equivalence reduces acceptance from 107/128 to 0/128 using a new harvest.

  • The evaluation includes held-out prompts, multiple output lengths, two detectors, and automated quality checks. The authors clearly acknowledge the limitations of the public implementation and the reported success metric.

Critical weaknesses
  • The theory lacks a nontrivial instantiation at realistic query budgets. Theorem 1 uses a stronger setting than standard undetectability [1], while Theorem 2 leaves flinkf_{\mathrm{link}} unspecified and uses a collision event that may become likely after many queries.

  • Formal spoofing success remains unverified. Appendix E.4 acknowledges that freshness cannot be checked against the complete harvest. The joint rates in Table 1 therefore do not establish success under the paper's own spoofing definition.

  • The strongest practical results rely on the implementation-specific bucket equivalence. In Table 8, the exact-context attack achieves 92.4% acceptance but a mean quality score of 1.54 and joint success of 0/512, limiting the empirical support for practical spoofing through general context reuse.

  • Related-work positioning and matched baselines are incomplete. The reuse mechanism deserves comparison with SEEK [3], spoofing-artifact detection [4], De-mark [5]. Scrubbing attacks and security trade-offs [6][7] provide further context. A controlled comparison with existing method [2] would help isolate the benefit of pooling.

  • The draft needs proofreading. “Sections 5 and 6 develops” should use “develop” (p. 2), and “The authors later application” needs a possessive (p. 7). The reference to Appendix F.4 for the λm\lambda_m and τ\tau ablations also appears incorrect: that appendix studies δ\delta.

[1] Miranda Christ, Sam Gunn, and Or Zamir. Undetectable Watermarks for Language Models. COLT, 2024.

[2] Nikola Jovanović, Robin Staab, and Martin Vechev. Watermark Stealing in Large Language Models. ICML, 2024.

[3] Huanming Shen, Baizhou Huang, and Xiaojun Wan. Enhancing LLM Watermark Resilience Against Both Scrubbing and Spoofing Attacks. NeurIPS, 2025.

[4] Thibaud Gloaguen, Nikola Jovanović, Robin Staab, and Martin Vechev. Discovering Spoofing Attempts on Language Model Watermarks. ICML, 2025.

[5] Ruibo Chen, Yihan Wu, Junfeng Guo, and Heng Huang. De-mark: Watermark Removal in Large Language Models. ICML, 2025.

[6] Abdulrahman Diaa, Toluwani Aremu, and Nils Lukas. Optimizing Adaptive Attacks against Watermarks for Language Models. ICML, 2025.

[7] Hanlin Zhang et al. Watermarks in the Sand: Impossibility of Strong Watermarking for Language Models. ICML, 2024.

Questions for authors
  • Can the authors give a nontrivial bound on flinkf_{\mathrm{link}} for the evaluated SynthID configuration and characterize how ε\varepsilon grows with the query budget? Does the bound remain informative after accounting for initialization contexts and reuse across responses?

  • Can freshness be evaluated against the full harvested response set for the pooled attack? Please specify the closeness predicate and threshold, and report the fraction satisfying freshness, detection, and quality together.

  • Can the exact-context attack produce acceptable-quality spoofs against the patched implementation with a larger budget? How does the pooled attack perform under reference-model mismatch and domain shift? These results would help distinguish the impact of the implementation weakness from that of general randomness reuse.

Reviewer_YYivOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies security weaknesses of LLM text watermarking caused by hidden randomness reusing. The authors establish a theoretical connection between watermark spoofing and undetectability under an additional verification setting. They also derive security bounds that account for repeated contexts. In the experiment, they develop 2 black-box spoofing attacks against SynthID-Text. The attack success rates are high, especially when exploiting hash-bucket reuse in the public implementation. The paper also evaluates a possible mitigation.

Critical strengths

The paper connects formal security analysis with practical attacks on a real watermark implementation.

The attacks success rates are high, and the proposed mitigation is useful.

Critical weaknesses

The strongest attack relies on a specific public implementation. It limites the generality of the findings.

The evaluation relies on automated quality checks and does not fully verify output freshness.

Questions for authors

Can the authors test whether similar weaknesses exist in other watermark implementations?

Can the authors provide stronger freshness checks and human evaluations of spoofed text?

Reviewer_gXu1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies randomness reuse in context-dependent LLM watermarks through formal analysis and spoofing attacks on SynthID-Text. It first reduces spoofing to a verification-augmented indistinguishability game that permits one terminal detector query. It then develops a security bound separating fresh-seed sampling, repeated-context randomness reuse, and generation–verification linkage. The empirical contribution comprises an exact-context probability-ratio attack and a stronger bucket-pooled attack. The latter exploits publicly identifiable equivalence classes in the evaluated Hugging Face implementation, allowing observations from different context–token windows to be pooled. On 512 held-out prompts, the pooled attack achieves joint detector-acceptance and automated-quality rates of 82.6% and 88.7% for Gemma-2B and Gemma-9B at 160 tokens. Controls include alternative detectors, permuted bucket assignments, and patched hashing. The strongest results concern this implementation-specific equivalence, while formal freshness is not fully verified.

Critical strengths

1. Concrete security finding. The paper identifies a public hash-bucket equivalence in the evaluated Hugging Face implementation. This explains how watermark information can be pooled across different contexts without knowing the secret key, providing an actionable finding beyond simply demonstrating successful spoofing.

2. Strong attack performance. On 512 held-out prompts, the pooled attack achieves joint detection-and-quality rates of 82.6% and 88.7% on Gemma-2B and Gemma-9B. Evaluating text quality alongside detector acceptance makes these results more practically meaningful.

3. Careful empirical validation. Unseen-window tests, permuted-bucket controls, and alternative-detector evaluations support the attack mechanism. The patched-hashing experiment reduces acceptance from 107/128 to 0/128, helping establish the role of public bucket equivalence and demonstrating a concrete mitigation.

Critical weaknesses

1. Limited generality and novelty.

  • (a) The pooled attack relies on public hash-bucket equivalence in the evaluated Hugging Face implementation of SynthID-Text [4]. The patch experiment supports this explanation by blocking the attack after removing that equivalence. Although the authors acknowledge this limitation, the evidence establishes an implementation-specific vulnerability. Which conclusions extend to SynthID-Text implementations or other watermarks without this hashing structure?
  • (b) The exact-context attack provides limited evidence that context reuse alone enables high-quality spoofing. In Table 8, restricted decoding achieves 92.4% acceptance but a mean quality score of 1.54 and zero joint success; the quality-preserving variant has zero acceptance. Can the authors demonstrate successful, quality-preserving spoofing without relying on public bucket equivalence?
  • (c) Probability-ratio watermark stealing predates this work [2], and attacks on SynthID-Text have already been studied [3]. The main new attack component appears to be bucket pooling. A comparison with an existing stealing method under the same query budget and evaluation protocol is needed to establish its contribution. How much improvement does pooling provide under such a comparison?

2. Gaps between theory and evaluation.

  • (a) Theorem 1 concerns verification-augmented undetectability, which grants a terminal detector query unavailable in the standard definition [1]. Its conclusion therefore does not directly apply to ordinary undetectability. Theorem 2 also leaves flinkf_{\mathrm{link}} unspecified, so the bound may be vacuous even without context collisions. Can the authors instantiate this term for a relevant construction and obtain a non-trivial spoofing bound? A weaker upper bound under reuse should also be distinguished from a proof that reuse enables an attack.
  • (b) The collision event BnB_n includes repeated active contexts across responses, while Appendix E.1 specifies a shared zero-padded initialization. If the initial positions are watermarked and masking history resets between responses, two nonempty queries already imply Pr⁡[Bn]=1\Pr[B_n]=1. Excluding these positions from the estimator does not remove them from the security game. Are they inactive, excluded from BnB_n, or subject to persistent masking? Please clarify when the bound remains non-vacuous.
  • (c) The empirical joint metric measures detection and quality, but not the freshness required by the formal spoofing definition. Appendix E.4 states that complete harvested responses were not retained, and its partial check concerns the exact-context attack. Freshness of the main pooled outputs is therefore unverified. Can the authors specify a closeness predicate and report joint detection, quality, and freshness results against all observed generation outputs?

References [1] Miranda Christ, Sam Gunn, and Or Zamir. Undetectable Watermarks for Language Models. COLT, 2024, pp. 1125–1139.

[2] Nikola Jovanović, Robin Staab, and Martin Vechev. Watermark Stealing in Large Language Models. ICML, 2024, pp. 22570–22593.

[3] Nikola Jovanović, Thibaud Gloaguen, and Martin Vechev. Probing Google DeepMind’s SynthID-Text Watermark. ETH Zurich SRI Lab technical blog, December 20, 2024.

[4] Hugging Face Transformers. SynthID-Text implementation, version 5.1.0. Official source code.

Questions for authors

see the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.