acceptodds
Under review as a conference paper at ICLR 2027

Twenty-Five Samples to Blind a Guardrail: Stealthy Benign-Side Poisoning of Self-Adapting Safety Filters

Abstract

Safety filters increasingly learn from experience: self-evolving guard pipelines periodically retrain on incoming traffic, and commercial moderation services learn custom categories from the operator's own stream. Because human annotation cannot keep up with volume, these loops label traffic with cheap, weak models. We show that this practice creates an overlooked attack surface: an adversary whose only capability is submitting prompts can poison the loop. The attacker wraps harmful queries in innocuous-looking framings and submits them at volume; the defender's own labelers naturally mislabel a few as safe, and the next fine-tuning step teaches the guard to tolerate that framing style, opening a hole for genuinely harmful requests. Every poisoning label is produced by the defender's pipeline, so the attack is indistinguishable from ordinary traffic. Across five guards from four families and two harmful-query pools, one batch of twenty-five naturally mislabeled submissions collapses or severely erodes framed-request detection on every guard, benign false positives stay at zero on every set we test, and end-to-end attack success against served targets rises from negligible to material. The realistic full-stream window sharpens the operating principle: the mislabel rate, not the dose alone, is the security parameter. Professional-task framings hold the mislabel rate near half, and a few dozen of them blind the guard surgically while every regression metric stays almost unchanged. We contribute methods for both sides: professional-task farming (PTF) and geography-guided farming (GGF) for the attacker, and anchor-calibrated loss filtering (ACLF) for the defender, which removes all of the poison in our runs at a bounded and front-loaded cost in benign data.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM safety, data poisoning, adversarial attacks, weak supervision, guardrails

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DVyJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes benign-side adaptation poisoning. Attacker only submits harmful queries wrapped in benign framings. Defender’s weak-labeler committee naturally mislabels some as safe. These mislabeled rows enter periodic fine-tuning and teach guard to ignore this framing. With 25 poison rows on ShieldGemma-2B, framed-harmful detection drops sharply while benign false-positive rate stays zero. Authors test dose, stealth, blind transfer, generality across guards and pools, pacing, margin mechanism, and defenses. They propose PTF and GGF attack variants and ACLF defense. Validation uses open-weight guards, public query pools, 3 seeds, and a primary adaptation recipe with additional ablations.

Critical strengths

(1) The threat model is novel and minimal: attacker only submits prompts, and poison labels are not injected or flipped but come from defender’s own weak labelers. The permission matrix clearly separates this channel from prior poisoning. If true, it exposes a shared risk of self-adapting guardrails in the labeling loop, so it is consequential.

(2) The mechanism analysis is useful: decision margins cross the boundary while weight displacement remains relatively stable, and ACLF calibrates loss threshold on trusted anchors to remove poison rows. This links attack to a practical defense and gives a concrete lever for defenders. These strengths do make sense because they initate a new attack surface and a testable defense, not just one more jailbreak.

Critical weaknesses

1.The critical weakness is evidence does not support production-level claim enough. The headline dose-25 attack is dose-parameterized, but authors’ own full-stream window shows simple-framing attack is completely neutralized, and professional-task attack only partly survives. So main result may be upper bound, not realistic damage. Stealth claim also relies on two model-generated and model-certified 320-item benign sets, not standard public benchmarks, and only a small number of seeds with very large variance. Zero false positive and collapse numbers are not robust enough. Defense comparison is only in same small pipeline.

2.A secondary weakness is presentation: some figures have text/line overlap and are hard to read, and appendix is split into too many small sections, which makes paper fragmented. This is not main rejection reason but lowers quality. Also, the paper gives inconsistent statements about code availability, leaving it unclear whether the implementation can be inspected during review.

Questions for authors
  1. Can you report benign false-positive rates on standard public benign benchmarks, beyond the two self-generated 320-item sets, and provide more seeds and confidence intervals for the reported E2 drops?
  2. In the full-stream realistic window, what is the end-to-end practical damage of PTF? Since simple-framing is neutralized and dose-parameterized results are upper bounds, should the full-stream results, rather than dose-controlled results, be the headline?
  3. How sensitive are conclusions to the adaptation recipe (last two layers, 100 steps, lr 1e-5, 40 anchors, committee composition)? Would standard full fine-tuning close the channel, and how do the stronger-labeler results limit the generality claim?
2 comments
AuthorsOpenReview ↗

Response to Reviewer DVyJ: Public Benchmarks, More Seeds, and Recipe Sensitivity

Thank you. All results and text changes in this response are included in the revised PDF uploaded with this rebuttal (changes marked in blue).

Q1 (Benign FPR on public benchmarks; more seeds; CIs). Every headline arm is now scored on XSTest [1] and OR-Bench [2], and FPR is reported against the matched clean arm, because both the unadapted guard and the clean adaptation loop over-refuse public benign prompts (XSTest 0.136 and 0.399; OR-Bench 0.029 and 0.151). No attack arm rises above its matched clean arm; the rate falls, because poisoning makes the guard less strict:

armXSTest: clean to attackOR-Bench: clean to attack
dose-25 cliff0.392/0.428/0.376 to 0.000 (3/3)0.151 mean to 0.000 (3/3)
1k dilution (frozen)0.119 to 0.0010.019 to 0.000
full-stream PTF0.399 mean to 0.1520.151 to 0.030

On seeds, the dose-25 E2 drop over five seeds is 62.2±35.7pp (per seed 97.5/61.2/97.5/37.5/17.5; all at least 17.5pp, none with an FPR rise), and we report this spread openly.

Q2 (Full-stream damage of PTF; should full-stream be the headline?). Yes, the revision leads with the full-stream window and labels dose-controlled arms single-window upper bounds. Against matched clean arms the professional-task channel loses 42.6±22.1pp (per-seed attack 0.442/0.025/0.325 against clean 0.688/0.704/0.679; absolute five-seed E2_tvd 0.33±0.21), while plain and 12-family detection stay at 0.988 and at or above 0.963. Against the unadapted guard, the reduction is partial in some seeds. The simple-framing attack is fully stopped in the same window (E2 = 1.00, 3/3 seeds), and that null result is the main reason the mislabel rate, rather than dose, becomes the headline. A sweep at a fixed 200-submission quota locates where the window's defense disappears: E2_tvd moves from 0.727±0.085 at r_eff near 0.04 through 0.341±0.140 near 0.16 to a 0.23 to 0.27 plateau at or above 0.28.

Q3 (Recipe sensitivity; would full fine-tuning close the channel?). The paper already compares four mechanisms and sweeps the lr-by-pacing surface. W1 also noted that the defense comparison lives in the same pipeline, so we re-ran ACLF under full-parameter fine-tuning: it still removes all 25 poisons in 3 of 3 seeds (E2 1.0/1.0/1.0 against 0.0/0.075/0.0 unfiltered), so the defense is not tied to the last-two-layer recipe. The remaining arms:

recipe / committeeresult
acute dose-25, lr 5e-6 / 1e-5 / 2e-5E2 drop 53.4±17.0 / 81.7±27.5 / 84.6±16.3pp
chronic 5-round, lr 1e-5 / 2e-5drop 6.3±1.3pp (healed) / 25.9±17.6pp (slow window)
full-parameter FT, dose-25, 3 seedsE2 0.750 / 0.075 / 0.013 vs clean 1.0; mean drop 72.1±40.9pp
L_h+L_g committee (39.8% mislabel)E2 0.283±0.245; plain detection falls to 0.49
L_q+L_g committee (0% mislabel)channel closed (pool 0/1200)
ACLF under full-parameter FTremoves 25/25 poisons; E2 1.0 / 1.0 / 1.0

Standard full fine-tuning therefore does not close the channel (three seeds, so evidence rather than a final limit), and a committee with one strong member that catches close to 100% of the reachable framings is safe by design.

W2 (presentation), plus the claims-framing part of W1. Accepted. The upper-bound and seed-variance parts of W1 are answered in Q1 and Q2, and the defense-scope part in Q3. In the revised PDF the abstract leads with the full-stream result, the appendix is merged into four parts, and the code is provided in the Supplementary Material.

References

[1] Röttger et al., XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models, NAACL 2024.

[2] Cui et al., OR-Bench: An Over-Refusal Benchmark for Large Language Models, ICML 2025.

Thank you for the newly added experiments and analysis. However, my concern remains that the paper's novelty is not competitive enough compared to the ICLR bar. As a result, I will keep my original score.

Reviewer_C6BrOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This papers proposes a benign-side adaptation poisoning attack against self-adapting LLM safety guard. The attack only submits prompt and exploits the labeling error to misalign safeguard when the mislabeled data are used to train them. The authors show that 25 mislabeled samples can significantly degrade the harmful detection across multiple safeguard. The author also propose serval methods to improve attack efficiency and a defense based on loss filtering to mitigate the attack.

Critical strengths
  1. The threat model sounds very interesting and practical. Unlike prior poisoning attacks, the attacker doesn't control the training data or lobbers, making the threat model more relevant to real-world systems. 2.The author comprehensively evaluated multiple guard models, datasets, attacks, and defenses, and the provides useful insights into how weak labeling errors affect the robustness of continuously adapted safeguards.
Critical weaknesses
  1. The main evaluation does not fully reflect the proposed threat model, as it should include both successfully mislabeled poisoning samples and unsuccessful submissions into the training process. This is because attackers do not know which samples are mislabeled during the attack. Most results should be updated to align with the proposed threat model to better support the paper's claims.
  2. The evaluation is limited, as it only involves 240 benign samples, whereas the poisoning rate should be much lower and the number of benign samples should be significantly higher. The experiments do not realistically reflect the poisoning rate in real-world AI systems, which typically involves large volumes of data. This limits the significance of the claimed effectiveness of the 25-sample attack.
Questions for authors
  1. Why do the main experiments only include successfully mislabeled poison samples?
  2. How does the attack perform when the benign data has a much higher number like 1k or 10k? Does the 25-sample attack still work under these conditions?
4 comments
AuthorsOpenReview ↗

Response to Reviewer C6Br: The Full-Stream Protocol and Benign-Pool Scale

Thank you. We agree with both objections. All results and text changes in this response are included in the revised PDF uploaded with this rebuttal (changes marked in blue). The two experiments below address the two objections in turn: first the protocol, then the scale.

W1 / Q1 (Training should include failed submissions too, not only the mislabeled ones). We ran the dose-controlled arms to isolate the dose response, but they assume the attacker knows which rows got in, so they cannot carry the main claim alone. The revision therefore makes the full-stream window the main protocol, in which every submission trains the guard at its natural label with caught rows included, and keeps the dose arms only as labeled upper bounds. The headline numbers under the full-stream window are:

full-stream arm (200 submissions)result
simple framings only (mislabel 3.5%)E2 = 1.00 in 3/3 seeds; the attack is fully stopped by its own caught rows
professional-task only (mislabel about 50%)E2_tvd = 0.33±0.21 over 5 seeds; plain detection ≥0.988
professional-task vs matched clean arms−42.6±22.1pp on the professional-task channel

To locate where the channel opens, we also swept the mislabel rate itself by mixing the two framing types at a fixed quota of 200 submissions:

mislabel rate r_effE2_tvd (3 seeds)
≈0.040.727±0.085 (caught rows teach detection)
≈0.160.341±0.140
≥0.280.23 to 0.27 (damage persists)

Plain and 12-family detection stay at or above 0.963 at every mix. The headline is therefore updated as follows: dose alone is not enough to predict damage, the mislabel rate of the group the attacker can reach is the stronger predictor, and the channel closes where the labeling stops erring. Because the sweep changes r_eff through the framing mix, we report this boundary as an observation of the tested setup rather than a clean threshold.

W2 / Q2 (Does dose-25 survive at 1k or 10k benign?). We ran the requested experiment and report both recipes without favoring either. The setup uses benign traffic of 1,000 and 10,000 unique public benign prompts from OR-Bench-80k [1], a harder-than-normal benign pool with the same fixed draw for all seeds. The paper's 240 rows were 20 unique strings repeated, which we now state openly. Each window adds 40 caught-harmful anchors and the same 25 mainline poisons (rows = benign + 40 + 25). The table shows the E2 drop in points against the matched clean arm of the same scale, where larger values mean more damage:

benign scalepoison sharefrozen recipe (100 steps)epoch-scaled (one epoch)
240 (paper)8.2%62.2±35.7 (5 seeds)not run
1,0002.3%40.1±16.335.0±20.6
10,0000.25%5.4±13.5 (per-seed 15.0/11.3/−10.0)30.0±10.8 (23.8/42.5/23.7)

The reading is clear at 1k: the same 25 poisons cost 35 to 40 points under both recipes. At 10k the outcome depends on the recipe. Under the frozen recipe the poisons do not cause reliable damage, because 100 steps sample only 800 of the 10,065 rows and most poisons are never trained on. Under epoch-scaled training, where all rows are seen, they again cost 30.0±10.8 points. Fixed dose is therefore not enough; damage depends on whether training actually samples the poison rows, and the attacker's answer to dilution is more submissions (rough estimate, as in Sec. 4.1 of the paper: about 570 submissions for a 20-poison batch at a 3.5% mislabel rate).

In addition, every dilution arm was scored on two public benign sets, XSTest [2] and OR-Bench [1]. The attack arm's false-positive rate never rises above its matched clean arm; it drops (XSTest 0.119 to 0.001 at 1k-frozen, 3-seed means), because poisoning makes the guard less strict.

References

[1] Cui et al., OR-Bench: An Over-Refusal Benchmark for Large Language Models, ICML 2025.

[2] Röttger et al., XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models, NAACL 2024.

Thank you for the rebuttal and additional experiments. They help address some of my concerns. One remaining question is whether the authors could combine these two settings: evaluate the attack under the full-stream protocol with a larger benign pool (e.g., 10k samples). This would help establish whether the attack remains effective when both conditions are considered simultaneously. I appreciate the authors' efforts in addressing my concerns and running additional experiments. I am inclined to raise my score based on the rebuttal.

AuthorsOpenReview ↗

Follow-up to Reviewer C6Br: The Two Conditions Combined

Thank you for actively participating throughout the discussion and for your question. We ran the combination, three seeds for the frozen recipe and one for the epoch-scaled recipe. The result maps the boundary of the attack: at 10,000 benign prompts the attacker's submissions are too small a share of the window, and damage is not distinguishable from zero at either volume or recipe.

Setup. 10,000 committee-safe unique benign prompts from Alpaca (ordinary user instructions; the committee flags 0.4 percent of them), 40 caught-harmful anchors, and the professional-task submissions at their natural committee labels. Two submission volumes (200 and 880) under the frozen recipe, plus an epoch-scaled run; a matched clean arm per seed. Alpaca rather than OR-Bench, because OR-Bench is deliberately near-boundary: at 10k the clean loop itself erases professional-task detection without any attack (0.00 against 0.32 unadapted).

At 200 submissions. Paired damage, clean minus attack in E2_tvd, is not distinguishable from zero:

recipecleanattackpaired damage
frozen, seed 00.3000.0040.296
frozen, seed 10.3380.3330.005
frozen, seed 20.0080.083−0.075
epoch-scaled, seed 00.2580.354−0.096

At 880 submissions. Raising the volume does not reopen it either, and the sign of the paired damage tracks which supervision the window samples more of:

seedpoison rows sampledcaught rows sampledcleanattackpaired damage
028370.3000.671−0.371
136290.3380.2500.088
230350.0080.671−0.663

Why. At 10k the 100-step budget samples about 800 of 10,200 to 10,900 rows, and the window carries both poison rows (labeled safe) and caught-harmful siblings (labeled unsafe). At 200 submissions the sample holds 8 to 9 poison rows and 5 to 8 caught rows, at or below the lower edge of the 20 to 25 row cliff that the 240-row window crosses, and the paired damage averages near zero. At 880 submissions it holds 28 to 36 poison rows and 29 to 37 caught rows, and the sign of the paired damage tracks which of the two is larger in all three seeds; we report that tracking as a hypothesis from three seeds, not an established mechanism. The clean arm is unstable at this scale for the same sampling reason (0.300, 0.338, 0.008), and in the single epoch-scaled run we observe no damage.

What this means. The channel requires the attacker's submissions to be a non-trivial share of the window, as in the 240-row regime of the main text (200 of 480 rows, about 42 percent of traffic). At 10,000 benign prompts the attack operates outside its effective regime, and we now state this scope limit explicitly. We are grateful for the question: it forced a boundary we had not measured.

Thank you for running the additional experiments. I appreciate the authors' efforts in addressing my concerns and reporting the negative results.

I encourage the authors to soften some of the claims about the attack's effectiveness and practical applicability, and to incorporate these additional experiments and discussions of the attack's limitations into the revised paper.

Overall, I appreciate the work and the constructive discussion. I will raise my score to Weak Reject.

Reviewer_E6mqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines a poisoning attack against safety guards that periodically retrain on user traffic. The attacker can only submit prompts; they cannot alter labels, training data, or model weights. By framing harmful requests as benign tasks, the attacker exploits errors in the defender’s automated labeling committee. Prompts incorrectly labeled safe then enter the next training update and reduce the guard’s ability to detect similar harmful requests. The paper evaluates the attack across several guards and query pools, studies submission costs and adaptation mechanisms, and proposes defenses based on filtering and trusted training examples.

Critical strengths
  • The submission-only threat model is clearly defined.
  • The paper evaluates multiple guards and includes a full-stream experiment that tests the effect of training on all submitted prompts.
Critical weaknesses
  1. The paper does not provide sufficient evidence that deployed safety systems commonly retrain on untrusted user traffic using a labeling pipeline comparable to the authors’ constructed weak committee. This assumption is central to the claimed practical relevance.

  2. The novelty is unclear. Exploiting mislabeled attacker-controlled data is an established poisoning mechanism. The paper should explain more precisely what its submission-only setting contributes beyond adapting this idea to safety guards.

  3. The “25 samples” result assumes that only the prompts mislabeled as safe enter training. In practice, the attacker submits all the prompts and cannot remove the ones labeled harmful. When the authors train on the full stream, the ordinary-framing attack no longer works. The paper should give this result more weight.

  4. The attack depends on how often the labelers make mistakes. The paper shows that a stronger labeling setup stops the attack on AdvBench, although the attack still works on StrongREJECT. This makes it unclear how broadly the findings apply to systems with stronger safety labeling.

Questions for authors
  1. What evidence supports the assumed weak labeling and retraining loop as representative of deployed systems?
  2. What is the precise technical novelty relative to poisoning through naturally occurring label errors?
  3. Can the authors present the full-stream results and raw submission costs alongside the 25-sample result in the main claims?
  4. Under what labeler error rate and realistic submission limit does the attack remain effective?
1 comment
AuthorsOpenReview ↗

Response to Reviewer E6mq: From a Dose Effect to a Security-Parameter Map

Thank you. All results and text changes in this response are included in the revised PDF uploaded with this rebuttal (changes marked in blue); seed-level tables are in the revised appendix. Taken together, the new results turn the four weaknesses into one quantitative claim: the mislabel rate, not the dose, is the security parameter.

W1 / Q1 (Is the weak-labeling retraining loop representative of deployed systems?). Each layer is either documented in prior work or measured in our own pipeline, and the revision states the assumptions openly in Sec. 3. One production deployment is documented [1], and continual adaptation on observed attack traffic has been developed and evaluated [2]. Commercial moderation services learn operator-defined custom categories from the operator's own stream [7]. Weak model labeling at volume then follows from cost: humans cannot label all traffic, so weak-supervision majority voting is standard practice [3]. New runs further show that the conclusion does not depend on our exact committee:

labeling stackmislabel rateattack outcome
L_q+L_g (two models)0% (pool 0/1200)channel closed
main committee (paper)3 to 4%reference result
L_h+L_g (no small model)39.8%poisoning gets easier

In short, the way labels are produced, rather than our particular committee, is the main factor.

W2 / Q2 (What is the precise novelty vs. poisoning through natural label errors?). The difference lies in who controls the poison. In label-noise poisoning the attacker supplies the wrong labels or writes into the training set, as in BadJudge [4] and Rapid Poison [5]. In our setting the attacker only submits prompts, and the defender's own committee picks which rows train the guard, at its own error rate. The same labeling function both filters the traffic and builds the training set. This link produces our new findings: dose alone does not predict damage while the mislabel rate does (W3/W4); one fast batch beats slow dosing when each round resets anchors; and at low mislabel rates the caught rows themselves act as a defense. These are properties of the adaptation loop, not of label noise by itself.

W3 / Q3 (Present full-stream results and raw submission costs alongside the 25-sample result). We agree, and the revision makes the full-stream window, in which every submission trains at its natural label with caught rows included, a main-text result. Dose-controlled arms are labeled upper bounds for one window, since they assume the attacker knows which rows got in. Raw submission costs (530 to 940 for AdvBench [6]; about 50 for professional-task) now sit beside every dose number.

full-stream armresultmeaning
simple framings, 3.5% mislabelE2 = 1.00, 3/3 seedsfully blocked by caught rows
professional-task, about 50% mislabelE2_tvd = 0.33±0.21 (5 seeds), plain ≥0.988channel partly survives
same arm vs matched clean arms−42.6±22.1ppcost vs what the loop would deliver

The revised abstract already leads with this operating principle.

W4 / Q4 (Under what error rate and submission limit does the attack remain effective?). In this setup, damage persists once the mislabel rate on an attacker-reachable group passes roughly 15%.

r_effE2_tvd (3 seeds)reading
≈0.040.727±0.085caught rows teach detection
≈0.160.341±0.140sharp drop
≥0.280.23 to 0.27damage persists

Plain and 12-family detection stay at or above 0.963 at every point. One caveat remains: the sweep changes r_eff by mixing framing types, so the boundary is an observation of this setup rather than a clean threshold. At a per-source quota Q, obtaining about 20 poison rows requires r·Q ≥ 20, which gives roughly 570 submissions at a 3.5% mislabel rate and about 40 to 50 submissions at 50% (rough estimates, matching Sec. 4.1 of the paper).

References

[1] Ming et al., Yesterday's Shield, Today's Spear: A Self-Evolving Safety Guardrail in Production, arXiv 2026.

[2] Sood et al., COPA: Continual Preference Optimization for Adaptive Prompt Injection Defense, arXiv 2026.

[3] Ratner et al., Snorkel: Rapid Training Data Creation with Weak Supervision, VLDB 2017.

[4] Tong et al., BadJudge: Backdoor Vulnerabilities of LLM-as-a-Judge, arXiv 2025.

[5] Huang et al., Rapid Poison: Practical Poisoning Attacks Against the Rapid Response Framework, ICML 2026.

[6] Zou et al., Universal and Transferable Adversarial Attacks on Aligned Language Models, arXiv 2023.

[7] Microsoft, Azure AI Content Safety: Custom Categories (official documentation), Microsoft Learn 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion

Dear AC and reviewers,

We thank the reviewers and the AC for a substantive discussion. Most concerns were answered with new experiments; the rest, on deployment grounding and novelty, are addressed through revised framing and analysis. The paper was revised throughout, with all changes marked in blue.

What we added. Ten new experiment families: five seeds for the headline result, benign pools at 1,000 and 10,000 prompts, two public benign benchmarks (XSTest, OR-Bench), a mislabel-rate sweep, committee-composition and full-fine-tuning arms, ACLF under full fine-tuning, and the combined full-stream-plus-10k setting that Reviewer C6Br requested.

The three reviews asked one shared question: how much of the headline survives under realistic conditions. The revision answers it in three steps.

  1. The full-stream window is now the main protocol, and dose-controlled arms are labeled upper bounds. The dose-25 upper bound is a 62.2±35.7 point drop over five seeds (minimum 17.5); the cliff occurs in three of five seeds, and the paper now reports it as such. We also report the main negative result with full weight: simple framings are fully neutralized by their own caught siblings (detection returns to 1.00 in 3 of 3 seeds). This yields the principle the abstract now leads with: the mislabel rate, not the dose, is the security parameter.

  2. Statistics are stronger. On both public benchmarks, benign false positives never rise above the matched clean arm in any attack arm.

  3. Scope is explicit. Full fine-tuning does not close the channel. ACLF under full fine-tuning still removes all 25 poisons in three of three seeds. The combined 10k setting does not open the channel, because the attacker's share of the window becomes too small. We state this boundary as an explicit scope limit in the appendix.

Per reviewer.

Reviewer E6mq: deployment grounding is now an assumption hierarchy; novelty is delimited by the permission matrix; full-stream results and raw costs are in the main claims; the error-rate boundary is mapped.

Reviewer C6Br: both objections became headline experiments. We thank the reviewer for the positive follow-up assessment; the requested combination was run in the follow-up round with three seeds and is reported honestly as a negative result, now part of the paper's stated scope.

Reviewer DVyJ: public benchmarks, five-seed intervals, recipe sensitivity, consistent code statements, and the appendix consolidation into four parts are all in the revision; figure overlaps receive a cleanup pass at camera-ready.

The revised paper is stronger than the submission: the attack surface is mapped with its boundaries, and the defenses are quantified on both sides. Thank you again for the review period.

Authors of Submission 21278

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.