acceptodds
Under review as a conference paper at ICLR 2027

CERTIFYING GENERATIVE LLMs AGAINST SUFFIX-BASED JAILBREAKS

Abstract

We propose Discrete-Continuous Randomized Smoothing (DCRS), to our knowledge the first framework that certifies the safety of smoothed LLM responses against suffix-based jailbreaks while accounting for both discrete token changes and continuous embedding displacement. Existing certified defenses adapt -style (masking- or erasure-based) certificates from text classification to the safety setting: they bound the number of altered tokens but ignore how far those tokens move in embedding space, and extensive masking can discard useful input information. DCRS instead combines discrete token selection with Gaussian embedding perturbations and applies hierarchical randomized smoothing to derive an certified radius for changes to at most token positions. The smoothed decision remains safe when at most token positions change and their joint embedding displacement is below . We specialize this guarantee to suffix attacks by padding the reference prompt with dummy tokens and representing suffix insertion as fixed-length replacement. This construction yields a sufficient condition for certifying every vocabulary-token assignment to the suffix slots, connecting continuous embedding radii to discrete adversarial suffixes. Experiments demonstrate non-vacuous certificates for multi-token suffix replacements and evaluate empirical jailbreak resistance alongside benign-task utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Randomized Smoothing, jailbreak

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1KYjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DCRS, which certifies the safety of a smoothed model’s response, rather than of an input filter, against suffix jailbreaks. In each random trial, DCRS adds Gaussian noise to the embeddings of a subset of token positions, maps each noisy embedding back to the nearest vocabulary token, generates a response, and labels it SAFE or not with a refusal-prefix rule. Using hierarchical randomized smoothing, it derives an embedding-space radius for changes to a limited number of positions. Appending dummy tokens to the prompt turns any suffix into a replacement of these tokens, so if the radius exceeds the largest possible embedding change of the suffix, every suffix of that length is certified. On 98 AdvBench prompts with Llama-2-7B-Chat (noise scale 5, 90% of positions perturbed), all prompts are certified for suffixes up to three tokens and 46 for five tokens, with low attack success against GCG, PAIR, and AutoDAN and a 2–3.5 point drop on ARC and HellaSwag.

Critical strengths

The dummy-token construction is simple and clean. It turns suffix insertion into a fixed-position replacement and gives a guarantee over every suffix of a given length, independent of how the attack is built; it also extends naturally to suffixes combined with prompt edits (Appendix N).

The certificate is on the generated response, which is what matters for jailbreaks, and it is cheap: mapping noisy embeddings back to tokens avoids a second denoising model, so a trial takes 0.17 s, compared with 0.15 s without defense and 3.51 s for SelfDenoise (Table 2).

Critical weaknesses

My main concern is whether the smoothed model still reads the prompt at the main setting, which decides what the certificate means. In Llama-2-7B, any two token embeddings are less than 1.8 apart (Appendix N), while the noise has a standard deviation of 5 in every direction, so after mapping back, the perturbed tokens seem almost unrelated to the original ones, i.e., about 90% of the prompt is replaced. This fits two results in the paper: attack success stays near the undefended level at noise scale 0.1 and drops sharply from 0.2, roughly where the noise starts to change the nearest token (Figure 2, right); and with every position perturbed, the mean lower bound on the safe-response probability is still 0.991 (Appendix E). Because the refusal-prefix rule counts replies starting with “I apologize” or “I do not” as SAFE, a model that is simply confused by garbled input would also be certified, and the certificate would say little about how the harmful request is handled.

The utility and empirical results may not use the certification setting. Section 5.1 points to Appendix C for the utility protocol, but the noise scale, sampling ratio, votes, and scoring are not given there, and the setting behind the attack-success columns of Table 1 is not stated. Given the first concern, a 2–3.5 point utility drop at the certification setting would be surprising; all three results should be reported under one setting.

Questions for authors

At the main setting, how often is a perturbed position mapped back to its original token, and what fraction of responses to benign prompts are labeled SAFE under the same smoothing? If benign prompts are also mostly SAFE, the certificate mainly reflects loss of input information.

Minor (no response needed): Appendix F reports 96/98 and 27/98 for DCRS at noise scale 1, which match the “+” placeholder in Table 7 rather than the default “!” (97/98 and 12/98 in Table 3). In Appendix C.4, the “i can’t” pattern lacks the start-of-response anchor that the other patterns have.

2 comments

Prompt legibility and utility at the certification setting (W1, W2, Q1)

DCRS at low noise is more accurate than every baseline at matched perturbation rates, and only its certificate survives when a refusal must show that the model understood the request. As the reviewer suspected, the main setting (σ=5\sigma = 5) neither keeps the prompt nor retains the reported utility; Table 2 did not measure utility at this setting. Measuring utility and certification at different settings is common among randomized defenses: RanMASK reports its best certificates at a 70–80% masking rate but its clean accuracy at 30% [1], and SmoothLLM reports its benchmark utility at a 5% perturbation rate [2]. At matched perturbation rates, DCRS at σ≤0.2\sigma \le 0.2 is more accurate than every baseline on ARC-Easy [3] (table below).

ARC-Easy (%), perturbed60%70%80%90%100%
No defense71.071.071.071.071.0
SelfDenoise [4]51.046.340.738.0–
RS-Del [5]50.344.339.336.0–
RanMASK45.041.038.335.7–
SmoothLLM (swap)36.338.037.736.7–
SmoothLLM (char. deletion)35.332.733.335.0–
DCRS σ=0.1\sigma = 0.170.771.069.770.069.7
DCRS σ=0.15\sigma = 0.1567.066.063.764.762.7
DCRS σ=0.2\sigma = 0.261.759.058.759.355.3
DCRS σ=5\sigma = 5 (paper setting: 90%)46.042.039.737.734.0

The baselines have no 100% entry because masking, deleting, or replacing every token leaves no input.

Refusals to an unreadable input say little about “how the harmful request is handled”. We therefore count a response as safe only when it clearly refuses the original request, naming its topic or referring to “your request”. Under this criterion, every baseline in the utility table above certifies a radius of zero at every rate, while DCRS at σ≤0.2\sigma \le 0.2 keeps a small non-zero radius (table below).

MethodMedian certified radius (k=1k = 1, ρ\rho units)
SelfDenoise, RS-Del, RanMASK, SmoothLLM (both), 60–90%0
DCRS σ=0.1\sigma = 0.1, 90%0.18
DCRS σ=0.15\sigma = 0.15, 90%0.22
DCRS σ=0.2\sigma = 0.2, 90%0.06
DCRS σ=5\sigma = 5, 90%0

We see these low-noise certificates as a first step for certifying generated responses against suffix jailbreaks, a new and difficult setting that we and others can build on.

The revised paper compares utility and certified radius at matched settings: it replaces Table 2 with the matched-setting utility above, adds Figure 3 in Section 5.3 that plots accuracy against this certified radius for every σ\sigma, rate, and baseline, and states σ\sigma, r, and the number of votes for every column of Table 1.

The “i can't” pattern (Minor)

Appendix C.4. The implementation already anchors the “i can't” pattern at the start of the response; the table in Appendix C.4 omitted the anchor, and the revised appendix corrects it. Anchoring it changes none of the reported counts.

References

[1] Jiehang Zeng, Jianhan Xu, Xiaoqing Zheng, and Xuanjing Huang. Certified robustness to text adversarial attacks by randomized [MASK]. Computational Linguistics 49(2), 2023.

[2] Alexander Robey, Eric Wong, Hamed Hassani, and George J. Pappas. SmoothLLM: Defending large language models against jailbreaking attacks. TMLR 2025.

[3] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? Try ARC, the AI2 Reasoning Challenge. arXiv:1803.05457, 2018.

[4] Zhen Zhang, Guanhua Zhang, Bairu Hou, Wenqi Fan, Qing Li, Sijia Liu, Yang Zhang, and Shiyu Chang. Certified robustness for large language models with self-denoising. arXiv:2307.07171, 2023.

[5] Zhuoqun Huang, Neil G. Marchant, Keane Lucas, Lujo Bauer, Olga Ohrimenko, and Benjamin I. P. Rubinstein. RS-Del: Edit distance robustness certificates for sequence classifiers via randomized deletion. NeurIPS 2023.

Dear Reviewer 1KYj, As the discussion period is nearing its end, we would like to kindly follow up on our responses. We hope our clarifications have adequately addressed your concerns, and we would greatly appreciate it if they could be reflected in your final assessment, if you find them satisfactory. Thank you again for your time and consideration.

Reviewer_WMs6OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies certification of a judged safety decision for generative LLM outputs under suffix-based jailbreaks. It proposes Discrete-Continuous Randomized Smoothing, combining random token-position selection, Gaussian embedding perturbations, vocabulary projection, and dummy-token padding to certify bounded-length suffix replacements. Experiments report multi-token vocabulary-wide certificates, empirical attack success rates, utility, and cost.

Critical strengths

The paper makes the certificate apply to the complete generation-and-judging pipeline rather than an input-side safety classifier. In particular, Equations 3–6 place vocabulary projection, generation, and the deterministic judge inside the base decision rule, and Theorem 4.2 and Corollary 4.3 provide a certificate jointly constrained by the number of changed token positions and their embedding displacement. This is consequential because it makes explicit what random variable is being certified in a generative setting.

The suffix reduction is concrete and empirically non-vacuous. Section 4.2 and Corollary 4.4 show that padding with dummy tokens converts arbitrary suffix insertion into replacement in a fixed-length space, with the sufficient condition covering all vocabulary assignments. Under the reported configuration, Section 5.2 certifies all 98 prompts through length three and 46 through length five, which is strong evidence that the derived condition can be achieved at meaningful discrete budgets.

Critical weaknesses

The certified property is a refusal-prefix label, not semantic response safety, and this distinction substantially limits the headline safety interpretation. Section 3.1 defines SAFE solely through the fixed prefix rule in Appendix C.4; a completion is SAFE if it matches one of several regular expressions, while every other response is NOT-SAFE. As the authors note in Section 6, a refusal-like opening followed by harmful content can be counted as safe, and harmless non-refusals are counted as unsafe. This is not merely conservatism: it means the formal result certifies stability of a brittle syntactic proxy rather than safety of the generated content. Since the title, abstract, and core framing repeatedly refer to certifying LLM safety, the current evidence does not establish that stronger semantic claim. The authors should narrow the language consistently to “certified refusal-prefix behavior” or provide a substantive validation of judge precision on the actual randomized generations, especially false-safe cases involving refusal-like openings followed by harmful assistance.

The reported multi-length certificates do not retain the stated per-test 95% confidence when interpreted as a joint guarantee through suffix budge K. Each padded length uses a separate one-sided Clopper–Pearson lower bound with α=0.05\alpha = 0.05 (Section 5.1, Appendix C.1), but certification through K requires all K bounds to hold. The paper itself acknowledges in Section 6 that the joint confidence is only at least 1−Kα1-K\alpha;thus the headline length-three and length-five claims have at best a union-bound confidence of 85% and 75%, respectively, per prompt, before considering the 98-prompt aggregate statement. This does not invalidate the deterministic implication conditional on valid probability bounds, but it weakens the statistical meaning of the reported certified-accuracy numbers. Recomputing with Bonferroni-adjusted per-length confidence levels, or reporting the joint confidence prominently alongside every multi-budget result, would resolve the issue.

Questions for authors

For the claims in Section 5.2 that prompts are certified “through” suffix budget K, will you report Bonferroni-adjusted results or joint confidence values? In particular, how many prompts remain certified through K=3 and K=5 when each per-length bound is constructed at level α/K\alpha / K?

In Appendix C.4, have you audited sampled DCRS completions that satisfy a refusal-prefix rule for harmful content appearing after the prefix? A high measured false-safe rate would materially change my confidence in interpreting the certificate as a safety guarantee rather than a refusal-style guarantee.

For the GCG results in Table 1, were suffixes optimized against the randomized DCRS majority-vote defense itself, including its noise and projection distribution, or generated against an undefended target and then evaluated after the fact? This determines how strongly the empirical ASR results support resistance to adaptive suffix optimization.

2 comments

False-safe audit of the refusal-prefix label (W1, Q2)

The measured false-safe rate is zero: no “refusal-like opening followed by harmful content” appears in a semantic audit of complete generations (table below). We generated responses of up to 512 tokens for all AdvBench prompts [4] and labeled every prefix-safe response with three independent safety classifiers: WildGuard [1], gpt-oss-20b [2], and Llama Guard 3 [3].

Judge (AdvBench, 98 prompts ×\times 50 samples, 512 tokens)Harmful among prefix-safe responses
WildGuard0 / 4,831
gpt-oss-20b0 / 4,831
Llama Guard 30 / 4,831

We use the refusal prefix because the certificate is meant to show that no suffix can change the model’s response to a harmful request. Without a suffix, the aligned model refuses these requests, and a suffix jailbreak succeeds exactly when the response stops starting with a refusal, so the prefix judge detects this change directly. On these generations, the audit above shows that this syntactic proxy agrees with semantic safety: all three classifiers judge every prefix-safe response safe.

The revised paper adds this audit to Appendix C.7 as the validation of the refusal-prefix judge.

Joint confidence through K (W2, Q1)

With each per-length bound at level α/K\alpha/K, every prompt remains certified through K=3K = 3, and only the K=5K = 5 count changes (table below). For each K, we recomputed the one-sided Clopper–Pearson bounds from the same 2,000 trials at level α/K\alpha/K; by the union bound, the certificates through K then hold jointly with probability at least 0.95.

LevelK=1K = 1K=2K = 2K=3K = 3K=4K = 4K=5K = 5
α=0.05\alpha = 0.05 per length (paper)9898989046
α/K\alpha/K per length9898989040

We report per-length confidence in the paper because each suffix attack has a single length: a 3-token suffix, for example, is blocked whenever the 3-token certificate holds, which has probability at least 95% on its own. The revised Appendix C.1 states this and reports both counts (Table 3).

Adaptive suffix optimization (Q3)

The GCG [4] suffixes in Table 1 were generated against an undefended target and evaluated after the fact, so Table 1 measures transfer rather than resistance to adaptive suffix optimization.

In a first adaptive attack optimized against DCRS itself, no prompt’s majority vote flips (table below). We ran GCG on the expected loss over the DCRS noise and projection (250 steps, 128 candidates per step, top-k 256, 8 noise samples) at σ=0.3\sigma = 0.3 on all 98 AdvBench prompts.

AttackJailbroken prompts, DCRS σ=0.3\sigma = 0.3 (10-vote majority)
Adaptive GCG (optimized against DCRS)0 out of 98 AdvBench prompts

The revised Table 1 caption states that its suffixes are transferred.

References

[1] Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. WildGuard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of LLMs. NeurIPS 2024 Datasets and Benchmarks Track.

[2] OpenAI. gpt-oss-120b & gpt-oss-20b model card. arXiv:2508.10925, 2025.

[3] Llama Team, AI @ Meta. The Llama 3 herd of models. arXiv:2407.21783, 2024.

[4] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv:2307.15043, 2023.

Dear Reviewer WMs6, As the discussion period is nearing its end, we would like to kindly follow up on our responses. We hope our clarifications have adequately addressed your concerns, and we would greatly appreciate it if they could be reflected in your final assessment, if you find them satisfactory. Thank you again for your time and consideration.

Reviewer_5zfMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Discrete-Continuous Randomized Smoothing to certify robustness against bounded suffix changes. The method randomly selects token positions, adds Gaussian noise to their embeddings, and projects the results back to vocabulary tokens. The model then generates a response, which a fixed judge labels.

Critical strengths

The link between embedding robustness and suffix coverage is clear. The padding construction gives an explicit condition covering every token assignment within the suffix budget. This is stronger than testing only suffixes found by a particular attack algorithm.

The experiments study several important factors. The paper examines noise scale, sampling ratio, placeholder choice, and embedding geometry. It also extends the analysis to attacks that combine suffixes with prompt-token changes.

Critical weaknesses

The certificate concerns a limited safety label. The judge accepts refusal-like prefixes, even when a response begins with a refusal and later provides harmful content. The paper acknowledges this problem but does not measure it. The certificate also preserves the majority label, not the safety of every response. A semantic audit and a clear description of the returned response are needed to support the practical safety claim.

The benefit of discrete selection and the utility of the strongest configuration remain unclear. The strongest certificates use σ = 5 and 90% token sampling, but the utility table does not clearly identify a matching configuration. It is therefore unclear whether the strongest guarantees retain the reported benign-task accuracy.

Questions for authors

How often does the prefix judge label harmful responses as safe under independent semantic or human assessment?

Can you report certification, benign utility, and total cost for the same configurations, including Gaussian-only smoothing at σ = 5?

How are finite-vote decisions converted into returned responses? What joint confidence applies across all certified suffix lengths?

3 comments

Semantic audit of the prefix judge (W1, Q1)

No responses “begin with a refusal and later provide harmful content” in a semantic audit of complete generations (table below). We generated responses of up to 512 tokens for all AdvBench prompts [4] and labeled every prefix-safe response with three independent safety classifiers: WildGuard [1], gpt-oss-20b [2], and Llama Guard 3 [3].

Judge (AdvBench, 98 prompts ×\times 50 samples, 512 tokens)Harmful among prefix-safe responses
WildGuard0 / 4,831
gpt-oss-20b0 / 4,831
Llama Guard 30 / 4,831

Preserving “the majority label, not the safety of every response” is not a limitation of DCRS; it is what randomized smoothing certifies by definition [5]. The smoothed model is the deployed model, and its output on each input is the majority label over randomized copies, not any single internal sample. For each input, the smoothed model draws N randomized copies, generates and judges one response per copy, and takes the majority label. The sampled responses are not shown to the user; the model returns only the first one, in sampling order, whose label equals the majority label. The returned response therefore carries the label that the certificate protects against every suffix of length k.

The revised paper adds this audit (Appendix C.7) and the definition of the returned response (Section 3.1).

Returned response and joint confidence (Q3)

The returned response is the first sample that carries the majority label, as defined in our answer to W1 above.

The joint confidence across all suffix lengths through K is at least 95% when each length is tested at level α/K\alpha/K, and only the K=5K = 5 count drops under this correction (table below). A statement that covers every length through K at once needs all K bounds to hold together; with each bound at level α/K\alpha/K, the union bound gives this with probability at least 0.95. We recomputed every one-sided Clopper–Pearson bound at α/K\alpha/K from the same 2,000 trials.

LevelK=1K = 1K=2K = 2K=3K = 3K=4K = 4K=5K = 5
α=0.05\alpha = 0.05 per length (paper)9898989046
α/K\alpha/K per length9898989040

The paper reports per-length confidence because each suffix attack has a single length: a 3-token suffix, for example, is blocked whenever the 3-token certificate holds, which has probability at least 95% on its own. The revised Appendix C.1 states this and reports both counts (Table 3).

References

[1] Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. WildGuard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of LLMs. NeurIPS 2024 Datasets and Benchmarks Track.

[2] OpenAI. gpt-oss-120b & gpt-oss-20b model card. arXiv:2508.10925, 2025.

[3] Llama Team, AI @ Meta. The Llama 3 herd of models. arXiv:2407.21783, 2024.

[4] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv:2307.15043, 2023.

[5] Jeremy Cohen, Elan Rosenfeld, and J. Zico Kolter. Certified adversarial robustness via randomized smoothing. ICML 2019.

Utility and discrete selection (W2, Q2)

DCRS at low noise is more accurate than every baseline at matched perturbation rates, and only its certificate survives when a refusal must show that the model understood the request. As the reviewer suspected, Table 2 did not measure utility at the certification setting, and the strongest guarantees do not “retain the reported benign-task accuracy”. This gap is common among randomized defenses. RanMASK reports its best certificates at a 70–80% masking rate but its clean accuracy at 30% [1], and SmoothLLM reports its benchmark utility at a 5% perturbation rate [2]. At matched perturbation rates, DCRS at σ≤0.2\sigma \le 0.2 is more accurate than every baseline on ARC-Easy [3] (table below).

ARC-Easy (%), perturbed60%70%80%90%100%
No defense71.071.071.071.071.0
SelfDenoise [4]51.046.340.738.0–
RS-Del [5]50.344.339.336.0–
RanMASK45.041.038.335.7–
SmoothLLM (swap)36.338.037.736.7–
SmoothLLM (char. deletion)35.332.733.335.0–
DCRS σ=0.1\sigma = 0.170.771.069.770.069.7
DCRS σ=0.15\sigma = 0.1567.066.063.764.762.7
DCRS σ=0.2\sigma = 0.261.759.058.759.355.3
DCRS σ=5\sigma = 5 (paper setting: 90%)46.042.039.737.734.0

Discrete selection also protects utility. For each σ\sigma in the table, perturbing only a subset of positions keeps accuracy above perturbing all of them, which is Gaussian-only smoothing (100%). The baselines have no 100% entry because masking, deleting, or replacing every token leaves no input, whereas Gaussian-only smoothing can still map each noisy embedding back near the original token.

A stricter certificate that counts only refusals of the original request also favors DCRS at low noise. Here a response counts as safe only when it clearly refuses the original request, naming its topic or referring to “your request”. Under this criterion, every baseline in the utility table above certifies a radius of zero at every rate, while DCRS at σ≤0.2\sigma \le 0.2 keeps a small non-zero radius (table below).

MethodMedian certified radius (k=1k = 1, ρ\rho units)
SelfDenoise, RS-Del, RanMASK, SmoothLLM (both), 60–90%0
DCRS σ=0.1\sigma = 0.1, 90%0.18
DCRS σ=0.15\sigma = 0.15, 90%0.22
DCRS σ=0.2\sigma = 0.2, 90%0.06
DCRS σ=5\sigma = 5, 90%0

We see these low-noise certificates as a first step for certifying generated responses against suffix jailbreaks, a new and difficult setting that we and others can build on.

The revised paper compares utility and certified radius at matched settings: it replaces Table 2 with the matched-setting utility above and adds Figure 3 in Section 5.3 that plots accuracy against certified radius for every σ\sigma, rate, and baseline.

References

[1] Jiehang Zeng, Jianhan Xu, Xiaoqing Zheng, and Xuanjing Huang. Certified robustness to text adversarial attacks by randomized [MASK]. Computational Linguistics 49(2), 2023.

[2] Alexander Robey, Eric Wong, Hamed Hassani, and George J. Pappas. SmoothLLM: Defending large language models against jailbreaking attacks. TMLR 2025.

[3] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? Try ARC, the AI2 Reasoning Challenge. arXiv:1803.05457, 2018.

[4] Zhen Zhang, Guanhua Zhang, Bairu Hou, Wenqi Fan, Qing Li, Sijia Liu, Yang Zhang, and Shiyu Chang. Certified robustness for large language models with self-denoising. arXiv:2307.07171, 2023.

[5] Zhuoqun Huang, Neil G. Marchant, Keane Lucas, Lujo Bauer, Olga Ohrimenko, and Benjamin I. P. Rubinstein. RS-Del: Edit distance robustness certificates for sequence classifiers via randomized deletion. NeurIPS 2023.

Dear Reviewer 5zfM, As the discussion period is nearing its end, we would like to kindly follow up on our responses. We hope our clarifications have adequately addressed your concerns, and we would greatly appreciate it if they could be reflected in your final assessment, if you find them satisfactory. Thank you again for your time and consideration.

Comments on the paper

1 comment

Overview

We thank all reviewers for their constructive feedback. The reviewers (5zfM, WMs6, 1KYj) noted that the certificate is on the generated response rather than on an input filter (WMs6, 1KYj), and that the dummy-token construction covers every suffix of a given length independently of how the attack is built (5zfM, WMs6, 1KYj). They also noted that the guarantee extends to suffixes combined with prompt edits (5zfM, 1KYj) and comes at a low per-trial cost (1KYj). Their concerns centered on three questions: whether the prefix judge labels harmful responses as safe (5zfM, WMs6), whether the smoothed model still reads the prompt and retains the reported utility at the certification setting (5zfM, 1KYj), and what joint confidence applies across suffix lengths (5zfM, WMs6).

Responses to the main concerns

Prefix judge (5zfM, WMs6). The semantic audit provides the requested validation of judge precision: none of the 4,831 prefix-safe responses, generated up to 512 tokens, is judged harmful by any of three safety classifiers, so the measured false-safe rate is zero (Appendix C.7).

Utility at the certification setting (5zfM, 1KYj). At matched perturbation rates, DCRS at low noise is more accurate on ARC-Easy than every randomized baseline, and up to 6.4 points more accurate than Gaussian-only smoothing without ever falling below it (Table 2). At these same settings, it is also the only defense that keeps a non-zero certified radius, under a judge that counts only refusals of the original request as safe (Figure 3, Appendix C.6).

Joint confidence across suffix lengths (5zfM, WMs6). The certificates also hold jointly. With each length tested at level α/K\alpha/K, so that all lengths through K hold together with probability at least 95%, every prompt remains certified through three tokens and the four-token count is unchanged; only the five-token count moves, from 46 to 40 (Appendix C.1, Table 3).

Other points. Section 3.1 now defines the returned response: the smoothed model returns only a sampled response whose label equals the majority label (5zfM). The GCG suffixes in Table 1 are transferred, as the caption now states. The certificate does not depend on how a suffix is optimized, so it also covers adaptive attacks up to the certified length (WMs6). Table 1 now states the DCRS setting and the number of randomized samples, and the “i can't” pattern is already anchored in the implementation, so no reported count changes (1KYj).

Summary of revision

We have uploaded a revised PDF, with changes marked in red.

Main

  • Abstract, Section 1, Section 7. Replace “retained benign utility” with the matched-setting result (5zfM, 1KYj).
  • Section 3.1. Define the response returned by the smoothed model (5zfM).
  • Section 5.1, Table 1. State that the GCG suffixes are transferred, define ASR, and give the DCRS setting of every row (WMs6, 1KYj).
  • Section 5.3, Table 2. Replace Table 2 with ARC-Easy accuracy at matched perturbation rates, including Gaussian-only smoothing (5zfM, 1KYj).
  • Section 5.3, Figure 3. Add certification at matched utility settings, plotting certified radius against accuracy for DCRS and every baseline (5zfM, 1KYj).
  • Section 6. Explain why the judge uses the refusal prefix rather than a harmfulness label, and point to the semantic audit (5zfM, WMs6, 1KYj).

Appendix

  • C.1, Table 3. Discuss confidence across suffix lengths and report counts at level α/K\alpha/K (5zfM, WMs6).
  • C.3. State the number of randomized samples and the ASR definition (1KYj).
  • C.4. Correct the anchor of the “i can't” pattern (1KYj).
  • C.5. Add the utility protocol (1KYj).
  • C.6. Add the certification protocol at matched utility settings and the request-grounded judge (5zfM, 1KYj).
  • C.7, Table 4. Add the semantic audit of the refusal-prefix judge (5zfM, WMs6, 1KYj).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.