acceptodds
Under review as a conference paper at ICLR 2027

Control-Plane Prompt Injection: Forging Authority in LLM Content Moderation

Abstract

Large Language Models (LLMs) deployed in content moderation increasingly process multi-channel inputs where untrusted user data can disguise itself as system infrastructure. We formally define and systematically evaluates this attack as control-plane prompt injection (CPPI), a vulnerability where attackers forge audit caches, label assignments, or formatting constraints within body text or user comments to hijack the control authority of the target model without altering the system prompt. CPPI exploits ambiguity between data and control signals, and causes models to copy forged labels or execution directives embedded in ordinary content. Extensive experiments show that CPPI successfully disrupts six representative defense measures on five mainstream LLMs. Our study identifies control-content confusion as a fundamental vulnerability of modern LLM moderation systems. We suggest that future prompt injection research should move beyond instruction conflicts toward understanding how models assign authority to heterogeneous input sources.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Control-plane prompt injection, structural provenance collapse, task-grounded hybrid discovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_C4FKOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Control-Plane Prompt Injection (CPPI), an attack targeting LLM content moderation systems that exploits structural provenance collapse in multi-channel inputs. CPPI forges system-like artifacts within user content to trick models into treating untrusted data as authoritative control signals. The authors formalize this via an 8-tuple grammar, use a QLoRA-assisted hybrid search to generate a frozen benchmark (CPPI-6), and evaluate across five mainstream LLMs and six defenses. Results show CPPI significantly outperforms the BIPIA baseline, with all tested defenses failing to mitigate it effectively.

Critical strengths

The first major strength is the systematic formalization of the attack space. Rather than relying on ad-hoc prompt crafting, the authors provide a structured eight-tuple grammar and a reproducible discovery loop that transforms prompt injection research from artisanal exploitation into a methodical engineering discipline. This framework offers clear value for future security benchmarking and allows other researchers to extend the attack taxonomy in a principled manner. The second strength lies in the comprehensiveness and statistical rigor of the empirical evaluation. Testing across five models, six defenses, and two datasets with generalized estimating equations and bootstrap confidence intervals sets a high standard for adversarial machine learning research. The finding that multiple defenses effective against traditional prompt injection fail against this new attack variant provides actionable evidence for the security community about the fragility of current mitigation strategies.

Critical weaknesses

The primary weakness is limited novelty beyond the systematization of known phenomena. The core insight that serialized multi-field inputs lose provenance boundaries is a well-documented limitation of transformer architectures in the security literature. While the mathematical formalism adds structure, the paper does not uncover fundamentally new mechanistic understanding or propose architectural solutions. For a venue like ICLR, this contribution reads more as a thorough benchmark consolidation than a scientific advance pushing the frontier of knowledge.

The second weakness is the complete absence of constructive defense exploration. The paper exhaustively demonstrates that existing defenses fail but offers no new mitigation mechanisms, architectural modifications, or even proof-of-concept interventions. For a security paper at this level, identifying a critical vulnerability without exploring any path toward remediation limits practical impact and leaves the community with a well-characterized problem but no directional guidance for solutions.

Finally, on the unvalidated mechanism claims, the assertion that the attack works through data-to-control promotion via attention alignment is central to the theoretical contribution, yet no attention visualizations, probing experiments, or representation analyses are provided to support this claim. Without empirical grounding, this remains a plausible hypothesis rather than a validated mechanism, which weakens the technical soundness of the core argument.

Questions for authors
  1. Can you provide attention map visualizations or probing results demonstrating that forged control tokens activate internal representations similar to genuine system prompts and distinct from regular user content? Without such evidence, how should the community distinguish the proposed mechanism from simple pattern matching or instruction following?

  2. Have you tested this attack against moderation systems that use constrained decoding, JSON mode, or function calling to enforce output structure architecturally rather than semantically? If not, what evidence supports the generalizability of your claims about threatening modern large language model moderation deployments in production environments?

  3. Given that the core vulnerability stems from serialization-induced provenance loss, what concrete architectural or training-time interventions do you believe could address this root cause? Even preliminary ideas, negative results, or failed mitigation attempts would substantially strengthen the constructive value of this work.

5 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/2]

Thank you for recognizing the systematic attack formalization and the breadth of our evaluation. The revised manuscript directly addresses your three central requests: controlled internal measurements, evaluation under enforced output formats, and concrete interventions guided by the identified vulnerability. Together, these additions connect the attack formulation to measurable changes in representations, decisions, and robustness.

1. Scientific contribution: identifying and testing forged moderation state

CPPI identifies the adoption of attacker-authored moderation state as an operational failure of the trust boundary. The attacker places an apparent audit result, cached label, or parser-facing decision inside an ordinary body or comment field. The moderator continues performing its assigned classification task and can return a fully valid output, while deriving the decision from this forged state. The trusted policy, model weights, API configuration, and message roles remain fixed (Section 3.1).

The original CHASM experiment already indicates a consequential difference: under Direct-task Excision, CPPI retains 43.9% macro ASR, compared with 16.7% for the calibrated BIPIA baseline (Table 1).

The eight-factor grammar makes this failure experimentally testable: it separates authority cues, target assignments, semantic reframing, and output contracts from channel, position, language, and repetition. The resulting scientific contribution includes controlled attribution of the forged target field, internal representation measurements with matched controls, valid-but-incorrect decisions under constrained outputs, and targeted intervention results. These linked findings establish which components drive the vulnerability and which changes suppress it (the revised Appendices A.4.2, A.6, and A.7).

2. Direct evidence from attention, representations, and controlled interventions

The revised Appendix A.6.1 now provides the attention visualizations and probing experiments you requested. We evaluate the open-weight Qwen2.5-VL moderator under five conditions: clean input, genuine system-role placement as a diagnostic reference, forged control text inside user content, ordinary quotation of the same wording, and length-matched neutral text. Across 613 examples and six template shells, this yields 18390 condition records, with measurements at layers 0, 7, 15, and 27.

Figure A2 measures clean-centered representation changes, Δhv(ℓ)=hv(ℓ)−hclean(ℓ)\Delta h_v^{(\ell)}=h_v^{(\ell)}-h_{\mathrm{clean}}^{(\ell)}. At layer 27, the mean cosine alignment of the forged-control change with the genuine-system change is 0.7418, compared with 0.3313 for its alignment with the neutral-text change. At layer 15, the corresponding values are 0.5327 and 0.3602. This centering focuses the comparison on the change induced by each input condition.

A logistic probe trained on genuine-system versus ordinary-quotation representations gives the held-out forged condition a mean system-reference score of 0.0884, compared with 0.0003 for ordinary quotation at layer 27. Cross-validation groups every variant and template of the same original record together, and preprocessing is fitted within training folds. Figure A3 additionally visualizes per-head attention allocation to system, payload, and other token regions.

These internal measurements complement the controlled target-field intervention in Table A3: removing the target assignment reduces GPT-4o ASR from 51.54% to 10.77%, while retaining the carrier and surrounding text. Across all five moderators, target-active odds ratios range from 1.637 to 5.064, with every 95% confidence interval above 1. The same-wording quotation control, length-matched neutral control, and target-field intervention together test how the presentation and content of forged control claims affect model behavior. These experiments substantiate control-content-induced changes in representations and moderation decisions. Appendix A.6.1 formulates the specific key-query causal pathway as a mechanistic hypothesis.

AuthorsOpenReview ↗

Reply to Reviewer [2/2]

3. Constrained decoding, JSON Schema, and function/tool interfaces

The revised Appendix A.6.2 now evaluates all three interface classes.

  • Native constrained decoding: the open-weight moderator is restricted to labels 0 and 1. CPPI achieves 57.80% mean ASR with zero invalid labels. Among 400 clean-correct examples evaluated with six templates, 988/2400 template-example pairs become false negatives, giving a newly induced error rate of 41.17%.
  • Hosted JSON Schema requests: mean CPPI ASR is 44.54% on GPT-4o, 62.62% on Gemini, and 50.27% on GLM-5. These measurements evaluate the hosted schema-request interfaces. The directly verified native decoding constraint is the preceding open-weight condition.
  • Function/tool interfaces: the frozen multilingual-review template produces valid incorrect label arguments on 611/613 GPT-4o examples and 608/613 Gemini examples, with zero invalid outputs at these endpoints. It overturns 539/541 and 369/374 clean-correct decisions, respectively.

These experiments demonstrate the practical distinction between output validity and decision integrity. A deployment can successfully enforce its response structure while accepting a moderation decision induced by untrusted control claims.

4. Constructive interventions and architectural/training implications

The revised Appendix A.7 introduces and evaluates two concrete interventions: control-field sanitization and provenance-aware serialization. Both run before moderator inference and require no additional model call or weight update.

The sanitizer replaces matched binary control values with an untrusted-value marker while preserving surrounding text and promotional evidence. Its frozen rules treat both binary values symmetrically and use neither ground-truth labels nor per-example template identities at inference. Table A9 reports the following mean-ASR changes across the six frozen templates:

ModeratorBeforeAfter sanitizationClean accuracy beforetextClean accuracy after
GPT-4o23.60%7.99%77.28%77.41%
Gemini-2.5-Flash46.38%27.62%84.42%83.41%
GLM-539.75%19.71%80.71%80.71%
DeepSeek-V4-Pro32.16%18.98%80.78%80.31%
Qwen2.5-VL-7B + Round479.91%64.74%78.96%78.96%

These reductions span 13.18–20.04 percentage points, with clean-accuracy changes of −1.01 to +0.13 points on 1483 positive and negative examples. The newly induced error rate of GPT-4o also falls from 19.19% to 3.62%.

The second intervention combines explicit JSON serialization of untrusted fields with a trusted provenance reminder. It reduces mean ASR from 23.60% to 4.49% on GPT-4o, with clean accuracy changing from 77.28% to 75.72%, and from 39.75% to 18.24% on GLM-5, with clean accuracy increasing from 80.71% to 82.27%.

The complete provenance-aware serialization results in Table A9 are:

ModeratorMean ASR beforeMean ASR afterClean accuracy beforeClean accuracy after
GPT-4o23.60%4.49%77.28%75.72%
Gemini-2.5-Flash46.38%39.83%84.42%84.09%
GLM-539.75%18.24%80.71%82.27%
DeepSeek-V4-Pro32.16%24.12%80.78%80.04%
Qwen2.5-VL-7B + R479.91%80.04%78.96%72.15%

Provenance-aware serialization reduces mean ASR on all four API moderators by 6.55-21.51 percentage points. On Qwen + R4, CPPI retains 80.04% mean ASR, compared with 79.91% before intervention, while clean accuracy changes from 78.96% to 72.15%. Together with the 64.74% residual ASR after control-field sanitization, these results show substantial CPPI persistence on Qwen under both targeted interventions.

These are attack-aware interventions designed with knowledge of CPPI and evaluated against the frozen CPPI-6 suite. Their role is to measure the functional importance of forged control fields under targeted intervention. Residual mean ASR reaches 64.74% on Qwen after sanitization, showing substantial remaining attack effectiveness even under this targeted intervention. The results motivate two concrete next steps: keeping authenticated moderation state in a trusted channel separate from user-authored records, and training on matched trusted/untrusted presentations of identical control claims. Appendix A.7 identifies these architectural and training directions as prospective extensions of the experimentally evaluated input interventions.

AuthorsOpenReview ↗

Deadline for Discussion is Approaching in Less than 3 Days

Dear reviewer C4FK ,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Reviewer_F8uiOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates Control-Plane Prompt Injection (CPPI), an attack against LLM-based content moderation systems that process multi-channel user inputs. The authors argue that flattening structured input fields into a unified textual context can blur the distinction between untrusted content and authoritative control instructions, allowing attackers to manipulate moderation decisions without modifying system prompts or model parameters. The authors evaluate CPPI against five LLMs on the CHASM and WGNA datasets, comparing it with an adapted BIPIA baseline under several defense configurations. The experimental results indicate that CPPI achieves higher attack success rates than BIPIA in many settings, particularly when task-oriented filtering and input-isolation defenses are applied. Additional experiments analyze the effects of forged target labels, injection channels, payload placement, repetition, and language variations.

Critical strengths
  1. The paper addresses a meaningful vulnerability in LLM-based content moderation systems. In realistic applications, user-controlled content may contain text resembling administrative instructions, evaluation metadata, or output-format requirements. Investigating whether models mistakenly follow these instructions is relevant to the security and reliability of deployed LLM systems.

  2. Rather than relying on a single manually designed adversarial prompt, the authors organize attack construction into multiple dimensions, including authority forgery, target-label assignment, semantic reframing, and delivery characteristics. This provides a useful framework for systematically examining different forms of authority impersonation in untrusted content.

  3. The experiments include five target models and two different content moderation datasets. The evaluation also considers multiple defense configurations, providing evidence that the investigated attacks are not limited to a single model or input setting.

Critical weaknesses
  1. Although the paper introduces the concept of "control-plane prompt injection" and formalizes it using an eight-factor attack grammar, the proposed approach appears to be primarily a task-specific reformulation and integration of existing prompt injection techniques, rather than a fundamentally novel attack paradigm.

  2. The paper attributes CPPI to "structural provenance collapse" and a "data-to-control promotion" process, further suggesting that forged authority markers induce key-query attention alignment with patterns characteristic of privileged system instructions. However, the analysis primarily relies on standard attention equations and observed classification outcomes. The manuscript does not provide direct measurements of attention redistribution, representation similarity to system-level instructions, or causal interventions supporting this proposed mechanism.

  3. The overall presentation quality of the paper requires substantial improvement. The writing style and figures exhibit characteristics commonly associated with AI-generated content, raising concerns about the care and attention devoted to manuscript preparation. Moreover, the experimental results are primarily summarized in a single main-text table (Table 1), whose formatting is particularly poor and significantly affects readability. More importantly, the selection of baseline methods is extremely limited, making it difficult to adequately assess the effectiveness and competitiveness of the proposed approach against existing prompt injection techniques.

Questions for authors
  1. Could the authors clarify the fundamental distinction between CPPI and existing prompt injection approaches involving authority impersonation, fake system messages, fabricated execution states, or output-format manipulation? Beyond introducing a structured attack grammar, what new attack capability or mechanism does CPPI establish?

  2. Could the authors provide comparisons with a broader range of representative baseline methods? The current evaluation relies on a rather limited set of baselines, particularly BIPIA, which is insufficient to comprehensively establish the effectiveness and competitiveness of the proposed approach. Including more recent and stronger prompt injection methods would help better demonstrate the advantages and novelty of CPPI.

8 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/2]

We thank the reviewer for recognizing the practical relevance of the moderation setting and the value of the structured attack space. The revised manuscript adds a feedback-optimized TextGrad comparison, controlled internal measurements, constrained-output experiments, and targeted interventions. These additions substantiate the contribution through independently checkable behavioral, representational, and intervention evidence.

1. The contribution: forging moderation state in ordinary user content

CPPI identifies and tests a specific failure of moderation authority: user-authored claims about a completed review, a cached decision, or a required output value can become the basis of the moderator's decision. The trusted policy, model weights, API settings, and message-role boundaries remain fixed. The attacker provides text through the ordinary body or comment fields and presents the desired label as an existing moderation result.

Sections 3.1–3.2 formalize this attack through four control-forging factors—authority/state, target assignment, semantic reframing, and output contract—and four delivery factors. The defining attack objective is to substitute forged task-internal state for an evidence-based moderation judgment. A forged review can preserve the expected evidence-and-label format while replacing the decision that the task requires the model to compute. Apparent compliance with the moderation workflow becomes a means of controlling its outcome.

The original comparison also exposes the practical difference: under Direct-task Excision on CHASM, CPPI retains 43.9% macro ASR, compared with 16.7% for calibrated BIPIA (Table 1).

This formulation yields concrete experimental predictions: neutralizing the target assignment should weaken the attack; forged control text should produce different internal changes from ordinary quotation; and valid output formatting should coexist with an incorrect decision. The revised experiments test all three. In particular, native binary constrained decoding still permits 57.80% mean ASR with zero invalid labels, including 988 newly induced false negatives among 2400 clean-correct template–example pairs (the revised Appendix A.6.2). These findings establish the practical significance of forged moderation state and connect the grammar to measurable model behavior.

2. Stronger baseline: feedback-optimized TextGrad with explicit query accounting

The revised Appendix A.5.1 adds the requested optimization-based comparison. The recovered CPPI discovery history contains four rounds of 24 candidates evaluated on 20 examples, totaling 1920 completed moderator-feedback evaluations. The new GPT-4o TextGrad search uses 96 candidate slots and a fixed 20-example CHASM development pool, consumes 1920 target-call attempts, including two failed requests, and records 285 optimizer calls separately. Six candidates are frozen by development ranking, including the top-ranked neutral initialization.

The comparison therefore aligns the scale of moderator feedback and reports the additional optimization calls explicitly. The historical CPPI development union contains 48 examples, and the new TextGrad search uses a fixed 20-example pool. TextGrad uses body injection, and CPPI retains its frozen body, comment, and dual-channel carriers. Both complete suites are evaluated under the permitted user-content threat model.

For transfer, Webis-500 comprises the first 500 positive records under an outcome-independent SHA-256 ID ordering. The new search does not optimize on this panel, and CPPI transfers without rewriting. Under the same strict scoring rule, Table A7 reports:

ModeratorTextGrad mean ASRCPPI mean ASRTextGrad coverageCPPI coverageTextGrad NERCPPI NER
GPT-4o2.97%26.80%4.2%72.4%0.04%21.98%
Gemini-2.5-Flash32.03%58.07%42.0%93.6%13.90%45.23%
GLM-514.27%39.80%17.8%99.4%3.22%31.06%
DeepSeek-V4-Pro14.43%43.40%20.6%83.0%4.15%33.13%
Qwen2.5-VL-7B + R494.60%88.17%98.8%100.0%75.48%83.72%
AuthorsOpenReview ↗

Reply to Reviewer [2/2]

Coverage counts examples successfully attacked by at least one of the six templates. Invalid outputs and failed requests count as non-successes. NER averages attack success over each suite's own matched clean-correct examples. CPPI improves mean ASR by 23.83–28.97 percentage points on all four API moderators and achieves higher coverage on all five. On Qwen, TextGrad has higher aggregate ASR, but CPPI has higher newly induced error rate, 83.72% versus 75.48%, using each suite's matched clean run. The new-error rates also favor CPPI on every API moderator. The revised evaluation thus combines the original calibrated direct-instruction baseline with a feedback-driven optimizer and measures their transfer beyond the development dataset.

3. Direct internal measurements and functional interventions

The revised Appendix A.6.1 adds the representation measurements requested in the review. We evaluate clean input, genuine system placement, forged control text, ordinary quotation of the same wording, and length-matched neutral text. Six template shells and 613 examples produce 18390 condition records, measured at layers 0, 7, 15, and 27.

The analysis measures clean-centered changes, Δhv(ℓ)=hv(ℓ)−hclean(ℓ)\Delta h_v^{(\ell)}=h_v^{(\ell)}-h_{\mathrm{clean}}^{(\ell)}. At layer 27, the forged-control change has mean cosine alignment 0.7418 with the genuine-system change, compared with 0.3313 with the neutral change (Figure A2). A logistic probe trained on genuine-system versus ordinary-quotation representations assigns held-out forged text a mean system-reference score of 0.0884, compared with 0.0003 for ordinary quotation. Cross-validation groups all variants of each original record together, and preprocessing is fitted in training folds.

The functional evidence links these measurements to the manipulated content. The controlled target-field ablation reduces GPT-4o ASR from 51.54% to 10.77% (Appendix A.4.2, Table A3). The new control-field sanitizer preserves surrounding content while neutralizing explicit user-authored decision values. Across five moderators, this reduces mean ASR by 13.18–20.04 percentage points, with clean-accuracy changes between −1.01 and +0.13 points (Appendix A.7, Table A9). These are attack-aware interventions designed against CPPI's frozen templates. To summarize, the results support the functional role of forged control fields and their associated internal representation changes.

4. Presentation and accessibility of the evidence

The revision gives each scientific question a dedicated presentation. Table A7 and Figure A1 report optimized-baseline transfer; Table A8 reports paired error transitions, Figures A2–A3 show measured internal responses, and Table A9 and Figure A4 report targeted interventions alongside clean performance. These focused tables and figures make the supporting evidence accessible beyond the original comprehensive Table 1. The compact baseline comparison above also makes the principal comparative result directly readable in this response. We have revised the rest part of the paper where necessary to further improve readability.

Thank you for the detailed response and additional experiments. While I appreciate the authors' efforts, I remain unconvinced that the rebuttal sufficiently addresses my primary concerns.

  1. The constrained-decoding experiment reports a mean ASR of 57.80%, while 988 newly induced false negatives among 2,400 clean-correct pairs correspond to only 41.17%. Although these figures may be based on different denominators or aggregation procedures, the relationship between them is not clearly explained. I believe the authors should provide a more transparent explanation of the evaluation protocol and statistical calculations.

  2. The manuscript's overall presentation, particularly its writing, figure design, and table formatting, remains below my expectations for a research paper. The original figures and tables appear insufficiently polished, and the presentation exhibits characteristics commonly associated with AI-generated content. I strongly encourage the authors to substantially improve the writing, redesign the figures, and carefully revise the tables rather than merely redistributing the experimental results across different sections.

  3. The fundamental novelty of CPPI remains insufficiently established.

Therefore, I will maintain my original rating.

Reviewer_pFT3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors study control-plane prompt injection, in which attackers place fake review records, classification labels, or formatting instructions inside content processed by an LLM moderator. They use a Qwen-based attack generator with iterative fine-tuning and moderator feedback to develop six fixed attack templates. Experiments evaluate these attacks across five moderator models, two advertising-detection datasets, and several defense configurations.

Critical strengths

I liked the detailed evaluation across five moderator models and two datasets, together with the ablation study of attack components. For example, removing the fake classification label from the multilingual attack reduces attack success on GPT-4o from approximately 52% to 11%, showing that the explicit label contributes substantially to the attack.

Critical weaknesses

The adapted BIPIA baseline uses one fixed instruction, whereas CPPI benefits from iterative training and selection using moderator feedback. I would like a stronger optimization baseline, such as GEPA or TextGrad, evaluated under a comparable query budget.

The contribution of training the attack generator is unclear. Comparing against attacks generated by the original Qwen3-Instruct checkpoint, with the same strategy prompts, candidate budget, and selection procedure, would help establish the benefit of iterative fine-tuning.

The aggregate moderation results before and after injection do not distinguish newly induced errors from pre-existing mistakes. I would like to see how many examples change from correct to incorrect and vice versa, together with an analysis of those changes.

The attention-based explanation is not directly supported by the reported experiments. Attention visualizations and, ideally, interventions on the open-weight moderator would help test whether the proposed mechanism explains the observed behavior. Otherwise, the explanation should be presented as a hypothesis.

Questions for authors

How does CPPI compare with a feedback-driven prompt optimizer such as GEPA or TextGrad under a comparable query budget?

How much does iterative fine-tuning improve attack effectiveness over generating and selecting templates from the original Qwen3-Instruct checkpoint?

How many examples change from correct to incorrect and from incorrect to correct after injection? What distinguishes these groups?

Can you provide direct evidence for the proposed attention mechanism, beyond the behavioral effects of changing attack text?

5 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/2]

We thank the reviewer for identifying four concrete ways to strengthen the empirical case. The revised manuscript adds a budget-explicit TextGrad comparison, an original-generator comparison, paired clean-to-attacked transitions, and controlled measurements inside the open-weight moderator. These experiments show that the frozen CPPI suite transfers effectively, overturns previously correct decisions, and produces measurable changes associated with forged control authority.

1. Feedback-driven optimization under a comparable target-feedback budget

We now compare CPPI with TextGrad at the scale of 1920 target-feedback calls and evaluate both frozen six-template suites on a fixed cross-dataset panel (the revised Appendix A.5.1, Table A7, and Figure A1).

The recovered CPPI discovery logs contain four rounds of 24 candidates evaluated on 20 examples, totaling 1920 completed moderator-feedback evaluations. The added GPT-4o TextGrad search uses 96 candidate slots on a fixed 20-example CHASM development pool, an adversarial objective targeting label 0, and an optimization-direction check. Its accounting records 1920 target-call attempts, including two failed requests, plus 285 optimizer calls. Six candidates are frozen by development ranking before transfer evaluation.

The comparison unit is the complete frozen attack suite within the permitted user-content surface: TextGrad uses body injection, and CPPI retains its body, comment, and dual-channel carriers. Appendix A.5.1 reports the development composition: the historical CPPI rounds span 48 distinct examples, while the TextGrad search uses a fixed pool of 20. The target-feedback accounting and optimizer-call accounting are reported separately from generator training.

Webis-500 comprises the first 500 of 6041 positive records ordered by the SHA-256 hash of their IDs. This selection rule fixes the transfer panel independently of attack outcomes. The new search does not optimize on this panel. Both methods use the same strict scoring rule, with invalid responses and request failures counted as non-successes.

ModeratorTextGrad mean ASRCPPI mean ASRTextGrad six-template coverageCPPI six-template coverageTextGrad NERCPPI NER
GPT-4o2.97%26.80%4.2%72.4%0.04%21.98%
Gemini-2.5-Flash32.03%58.07%42.0%93.6%13.90%45.23%
GLM-514.27%39.80%17.8%99.4%3.22%31.06%
DeepSeek-V4-Pro14.43%43.40%20.6%83.0%4.15%33.13%
Qwen2.5-VL-7B + R494.60%88.17%98.8%100.0%75.48%83.72%

CPPI improves mean ASR by 23.83–28.97 percentage points across all four API moderators and achieves higher six-template coverage across all five moderators. The clean-conditioned new-error rate (NER, to be specified in Item 3 of this reply) is also higher for CPPI on every moderator: 21.98%, 45.23%, 31.06%, 33.13%, and 83.72%, compared with TextGrad's 0.04%, 13.90%, 3.22%, 4.15%, and 75.48%, respectively, using each suite's matched clean run. These results establish a substantial transfer advantage against an explicitly feedback-optimized baseline.

2. Contribution of the iterative discovery workflow

The revision directly evaluates generation from the original Qwen3-30B-A3B-Instruct-2507 checkpoint using the same eight strategy families, three candidate slots per family, and a shared 20-example development panel (the revised Appendix A.5.2).

Across three evaluation seeds, the archived final multilingual-review template succeeds on 8/20, 8/20, and 9/20 examples. The development-selected original-generator candidates succeed on 2/20, 5/20, and 4/20. The corresponding mean ASRs are 41.67% and 18.33%, a 23.33-percentage-point improvement. The original-generator candidates subsequently achieve 14.2%, 22.3%, and 30.0% on the 613-example panel.

This result quantifies the contribution of the complete iterative adaptation, moderator-feedback, and historical-selection workflow. QLoRA adaptation is one component of that workflow, together with execution replay, validity screening, and development selection. We present the measured gain at this workflow level, consistent with the construction described in Section 3.3 and Appendix A.3.3. The attack-side Qwen3 generator and the evaluated Qwen2.5-VL moderator have distinct roles, as specified in Table A1.

AuthorsOpenReview ↗

Reply to Reviewer [2/2]

3. Newly induced errors and reverse transitions

The revised Table A8 now tracks predictions on the same records before and after injection, separating correct-to-wrong transitions from wrong-to-correct transitions. The following entries are representative individual-template endpoints. T1 is Audit Cache Override and T6 is Multilingual Review Note.

ModeratorTemplateCorrect → wrongRate among clean-correct examplesWrong → correct
GPT-4oT6300/57552.17%0
GeminiT1415/49484.01%0
GLM-5T1487/50297.01%0
DeepSeekT1224/52342.83%1
Qwen + R4T6314/32696.32%0

For GPT-4o, the 337 valid false-negative outputs comprise 300 newly induced errors and 37 persistent errors. The original promotional evidence remains in the record. The new error occurs after the forged review state is introduced. The newly induced group consists of records whose promotional evidence the clean moderator had already classified correctly. The persistent-error group consists of records already missed before injection. At these endpoints, the near absence of reverse transitions shows a strongly directional effect toward the attacker's requested label 0.

To make this distinction systematic across suites, Appendix A.5 introduces the new-error rate (NER). For the positive attack cohort and a suite of KK templates,

NER=1K∣C∣,∑i∈C;∑k=1KI[y^ik=0],C=i:y^iclean=1,\mathrm{NER}=\frac{1}{K|C|}\\,\sum_{i\in C}\\;\sum_{k=1}^{K}\mathbb{I}[\hat y_{ik}=0],\qquad C=\\{i:\hat y_i^{\mathrm{clean}}=1\\},

where CC contains the examples classified correctly in the corresponding clean run. The higher CPPI NERs reported above in Item 1 show that its transfer advantage includes newly created moderation failures.

4. Direct internal evidence and intervention results

The revised Appendix A.6.1 adds controlled attention visualizations and representation measurements for 18390 condition records: 613 examples ×\times six template shells ×\times five conditions. These conditions are clean input, genuine system-role placement, forged control text in user content, ordinary quotation of the same wording, and length-matched neutral text. Measurements cover layer indices 0, 7, 15, and 27.

Figure A2 compares clean-centered hidden-state changes. At layer 27, the forged-control change has mean cosine alignment 0.7418 with the genuine-system change, compared with 0.3313 with the neutral change. The layer-15 values are 0.5327 and 0.3602. A logistic probe trained on genuine-system versus ordinary-quotation representations, with five-fold cross-validation grouped by original content, assigns held-out forged inputs a mean system-reference score of 0.0884, compared with 0.0003 for ordinary quotations.

The intervention results establish the functional importance of the forged decision fields. Removing the target label reduces GPT-4o ASR from approximately 52% to 11%. The added targeted control-field sanitizer reduces mean ASR by 13.18–20.04 percentage points across all five moderators, with matched clean-accuracy changes between −1.01 and +0.13 percentage points (the revised Appendix A.7, Table A9). This intervention was designed with knowledge of the frozen CPPI templates and selectively neutralizes binary decision assignments while preserving the surrounding content.

The selected-head intervention changes the forged target rate by approximately +0.054 percentage points. The empirical mechanism is supported by measured representation shifts, controlled provenance comparisons, and decision-field interventions. Appendix A.6.1 presents key–query alignment as a mechanistic hypothesis for further testing.

AuthorsOpenReview ↗

Deadline for Discussion is Approaching in Less than 3 Days

Dear reviewer pFT3,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.