acceptodds
Under review as a conference paper at ICLR 2027

Implicit Ranking Leakage in Black-Box RAG: Retrieval Stability-Guided Knowledge Corruption Attacks

Abstract

Retrieval-Augmented Generation (RAG) is widely used in large language model systems. Recent studies show that RAG is vulnerable to knowledge corruption attacks (KCAs), where adversaries observe top-k retrieved documents to approximate the retriever and manipulate downstream generation. Existing attacks mainly use top-k membership as binary supervision and treat all retrieved documents as equally important. This ignores hidden ranking information and underestimates the risk exposed by retrieval behavior. In this paper, we propose Ranking Vulnerability-enhanced Knowledge Corruption Attack (RV-KCA), a black-box method that infers relative document retrieval stability from top- membership and uses it as pairwise supervision to conduct stronger knowledge corruption attacks. First, RV-KCA applies semantics-preserving query perturbations and estimates retrieval stability from how often each document remains in the top- set. Our theoretical analysis shows a monotonic relationship between retention probability and the original retrieval score under a shared local noise model. It also characterizes sampling uncertainty in the stability estimates, which serve as noisy evidence of relative document ordering. Second, RV-KCA selects document pairs with clear stability differences and uses them to train a margin-aware surrogate retriever. The surrogate model learns to assign higher scores to more stable documents while preserving membership supervision information derived from the observed top- sets. Finally, RV-KCA optimizes query-side triggers on the surrogate to give the injected document a ranking advantage over estimated boundary competitors across query variations. Experiments show that RV-KCA achieves higher surrogate fidelity and overall attack success rates than existing black-box baselines. It also reaches a fixed surrogate fidelity target with fewer target-retriever calls. These results demonstrate that retrieval stability exposes useful ranking information for knowledge corruption attacks even when similarity scores and explicit ranks are hidden.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
RAG, Knowledge Corruption Attacks, Stability

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_dDnsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper infers relative document preferences from retrieval stability under query perturbations. RV-KCA uses these preferences to train a surrogate retriever and optimize query triggers that promote poisoned documents.

Critical strengths
  1. The paper identifies useful information leakage from unordered retrieval results. Hiding explicit ranks therefore leaves a meaningful attack surface.
  2. The equal-observation controls help isolate the contribution of pairwise supervision. The shuffled-pair experiment provides an additional check on the inferred preferences.
  3. The attack-side ablation separates surrogate training from boundary selection. This gives the end-to-end improvement a clearer explanation.
Critical weaknesses
  1. The main comparison uses a query-side adaptation of FlippedRAG [1], whose published attack modifies documents. Could the stability-trained surrogate also improve the original document-only attack with the user query fixed?
  2. Retention can reflect a document’s sensitivity to the chosen rewrites as well as its original rank. Do the inferred preferences remain consistent under a different perturbation generator?
  3. Boundary competitors are selected before the trigger changes the query. Why should these documents remain the relevant competitors after optimization? Please compare the cached boundary set with periodically refreshed competitors.
  4. Can the query trigger directly influence the generator’s answer? Please repeat generation with the same retrieved passages while removing the trigger from the generator input.
  5. The target systems use exact dense retrieval without reranking. Does the inferred stability signal remain useful when a cross-encoder determines which passages reach the generator? Please test the complete pipeline with probing performed through its final unordered output interface.

[1] FlippedRAG: Black-Box Opinion Manipulation Adversarial Attacks to Retrieval-Augmented Generation Models.

Questions for authors

See Weaknesses.

12 comments
AuthorsOpenReview ↗

Response for Question 1

Comment 1:The main comparison uses a query-side adaptation of FlippedRAG [1], whose published attack modifies documents. Could the stability-trained surrogate also improve the original document-only attack with the user query fixed?

Response 1: Thank you for bringing this to our attention. We have added an experiment that applies stability supervision to FlippedRAG’s document-only attack while keeping the user query fixed. We preserve its document-side trigger placement and search objective, using our fact-substitution task. We use Qwen3-Embedding-0.6B with HNSW and Qwen3-Reranker-0.6B, paired with Qwen3-32B on NQ and MS-MARCO.

All variants share the surrogate architecture and cached training observations. The evaluation instances and document-search budget are also matched. Only the surrogate supervision changes. Query perturbations are confined to surrogate construction, while document optimization and attack execution retain the original user query. We additionally include a control that shuffles stability labels within each query.

Table 1. Stability supervision for FlippedRAG with fixed user queries.

CorpusSurrogate supervisionHit@10ASR
NQMembership only45.98 ± 1.0420.73 ± 1.07
NQMembership + shuffled pairs46.31 ± 1.1820.88 ± 0.91
NQMembership + stability pairs53.84 ± 0.9724.16 ± 0.96
MS MARCOMembership only48.31 ± 0.8920.84 ± 0.93
MS MARCOMembership + shuffled pairs47.96 ± 1.0920.59 ± 1.06
MS MARCOMembership + stability pairs54.92 ± 1.2123.49 ± 0.88

Stability supervision increases Hit@10 by 7.86 percentage points on NQ and 6.61 percentage points on MS MARCO. These retrieval improvements are accompanied by ASR gains of 3.43 and 2.65 percentage points, respectively. Therefore, the benefit reaches the final generated answer even when the attacker modifies only the document. The shuffled-pair control remains close to membership-only training on both datasets. This supports the contribution of informative document preferences beyond simply adding a pairwise training objective. Since the attack procedure and observations are matched, the comparison isolates the value of stability supervision within FlippedRAG’s document-side attack.

These results show that our surrogate-training contribution also benefits attacks with immutable user queries. The recovered ranking information can guide document edits while the original query remains unchanged throughout attack execution.

AuthorsOpenReview ↗

Response for Question 2

Comment 2:Retention can reflect a document’s sensitivity to the chosen rewrites as well as its original rank. Do the inferred preferences remain consistent under a different perturbation generator?

Response 2: We appreciate your comment. We have added a comparison across perturbation generators and clarified how document-specific rewrite sensitivity affects the interpretation of retention. We compare the original generator with Llama-3-8B-Instruct and Mistral-7B-Instruct. The evaluation uses NQ and MS MARCO with Qwen3-Embedding-0.6B retrieval and Qwen3-Reranker-0.6B reranking. Qwen3-32B remains the answer generator. Both alternatives follow the same rewrite families under the same semantic filter and probe budget. Only the generator used for surrogate training changes. Attack-stage probes are held fixed.

Direction agreement measures whether pairs selected by both generators receive the same preference. Joint coverage reports their fraction of all document pairs in the original top-10 set. The original-generator row compares two independent samples from that generator, providing a reference for variability under the fixed probe budget. Target ranks are used only to evaluate pair accuracy.

Table 2. Preference consistency across perturbation generators.

CorpusPerturbation generatorPair accuracy / coverageDirection agreementJoint coverageRV-KCA ASR
NQOriginal72.46 / 58.7386.4145.2625.68 ± 0.85
NQLlama-3-8B-Instruct73.08 / 56.9182.7342.1825.36 ± 1.12
NQMistral-7B-Instruct-v0.271.17 / 60.2480.9639.8724.82 ± 0.98
MS MARCOOriginal71.23 / 57.8684.7643.5923.92 ± 1.17
MS MARCOLlama-3-8B-Instruct70.64 / 58.4781.3840.8623.31 ± 1.04
MS MARCOMistral-7B-Instruct-v0.271.88 / 55.9379.8438.5224.08 ± 0.93

Across the alternative generators, 79.84–82.73% of shared preferences retain their direction, compared with 84.76–86.41% under independent resampling from the original generator. These cross-generator comparisons cover 38.52–42.18% of all original-query pairs. Pair accuracy also remains at 70.64–73.08%, showing that consistency is accompanied by alignment with the original target ordering. Changing the training probe generator alters mean ASR by at most 0.86 percentage points.

We have clarified that retention is conditioned on the perturbation distribution. Appendix B.1.5 accounts for document-specific sensitivity by requiring retention differences to exceed the uncertainty introduced by that sensitivity. The added experiments support a consistent interpretation of our contribution. The inferred preferences provide useful partial ranking information across the tested generators, while the selected pair sets can vary with the rewrites.

AuthorsOpenReview ↗

Response for Question 3

Comment 3:Boundary competitors are selected before the trigger changes the query. Why should these documents remain the relevant competitors after optimization? Please compare the cached boundary set with periodically refreshed competitors.

Response 3: We appreciate your thoughtful review. We have added a comparison between cached and periodically refreshed competitors, including a cached control with the same probing budget. We use the same HNSW pipeline on NQ and MS MARCO with Qwen3-Embedding retrieval and Qwen3-Reranker reranking. Qwen3-32B remains the answer generator. At each refresh, we append the current trigger to the original query and its four cached rewrites. We then reselect the three least-retained documents from the current top-10 set, excluding the injected passage. This uses the same unordered retrieval interface and keeps the surrogate frozen.

Table 3. Cached and refreshed boundary competitors.

CorpusCompetitor selectionMaximum probe calls per queryHit@10RV-KCA ASR
NQCached557.1125.68 ± 0.85
NQCached with more initial probes2058.4325.94 ± 1.08
NQRefresh every 20 updates1058.1726.63 ± 0.96
NQRefresh every 10 updates2060.3627.16 ± 1.12
MS MARCOCached554.7623.92 ± 1.17
MS MARCOCached with more initial probes2056.1124.29 ± 0.93
MS MARCORefresh every 20 updates1055.8324.62 ± 1.04
MS MARCORefresh every 10 updates2057.2424.88 ± 1.09

Refreshing every 10 updates increases mean ASR by 1.48 percentage points on NQ and 0.96 percentage points on MS MARCO over the default cached set. It permits up to 20 target probes per query, compared with five for caching. Against the cached control with the same probe allowance, the gains are 1.22 and 0.59 percentage points, respectively. This comparison supports the value of updating competitors beyond simply collecting more initial observations.

The default cached set nevertheless achieves ASR close to the refreshed variants with a smaller probing budget. These results support its use as a local approximation during optimization. We have clarified this role in Section 3.4 and included periodic refreshing as an option when additional target queries are available.

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

This paper study on black box knowledge corruption attack on RAG system, where the threat setting is that the retriever only returns an unordered topk documents and hiding both the affinity score and the rank order, so the attacker cannot see which doc is ranked higher. The paper's main idea is that the hidden ranking can be revealed by the stability of retrieval, which a document that is ranked high can sit farther above the topk cutoff, so it stays in the topk return more often when the user rewrite the queries manytimes. Based on this, the paper propose method named RV-KCA that has 3 steps including: rewrite each query several time to estimate the document's retantion frequency to represent its stability, then they make the gap into pairwise supervision and tune a dual encoder surrogate retriever with both membership loss and pairwise ranking loss, and finally picks the least stable docs as the boundary competitors and run beam search on the surrogate to optimize a short query trigger that pushes the inject docs above them. Experiment use 3 retriever and 3 models on NQ and TriviaQA, and the result shows it achieve the highest ASR in 8 model-data pairs.

Critical strengths

The paper is well-written and easy to understand, and the idea is interesting, clean and well-motivated, and it is reasonable that treat topk membership as a binary signal does throw away the ranking inside the retrieved docs, and recovering that ranking from how often a document survives query rewrites is a simple and intuitive way to extract more from the same black-box interface, the athor also connect this idea to the method design. Also, The paper is honest about the limits of its own theory that the finite-sample bound does not certify any pair at the budget it actually uses, as such kind of self scrutiny is easy to leave out, and this from my perspective make the claim and result easier to trust.

Critical weaknesses

Although the idea of recovering hidden ranking from top-k stability is interesting, and the paper is careful about what its theory does and does not show, I have two major concerns about whether this attack is meaningful in a realistic deployment, and both point to evaluations that are missing.

W1. The probing is a bit too expensive, highly conspicuous, and defeated by a simple adaptive defense.

The whole attack is on sending many near identical rewrites of the same query and also watching which documents stay in the topk. From the defender perpective this is quite unusual access pattern, and the paper never asks whether a deployed system would notice it, and it is easy to make an adaptive defense for this. Also the cost makes this worse rather than better, as the paper advertises efficiency, but it only saves on the surrogate-construction calls. By the accounting in appendix C.8, probing the 2k evaluation queris after inject already cost 10k extra target call, and together with the 15k construction calls one dataset-retriever pair needs about 25k calls, so the reported savings are relative to the construction block alone but not to this total.

The deeper problem is that the probing signal is trivial to suppress, because a defender who cache result for semantically near-duplicate queries, or who adds a small amount of randomness to tie-breaking or to the retrieval scores near the topk cutoff and then directly destroys the retantion frequency signal that the attack depends on, while almost not affecting normal user. The paper own theory makes this clear, as the method already relies on the empirical threshold delta, and a defender only has to push the retantion gaps below this threshold. The paper evaluates against no defense at all, I think this the major concern that need to be address. So I have two question here:

Q1. How many extra queries in total does the full attack issue per target query? how detectable is this volume of near-duplicate queries under a basic query-rate or query-similarity monitor?

Q2. How does the attack degrade under simple cheap defenses? for example near-duplicate query caching, or small random noise added to tie-breaking or to scores near the topk cutoff?

W2. The gain come almost entirely from retrieval, and the end-to-end effect on the LLM is both weaker and untested against simple defenses.

Also, from my perspective, for a knowledge corruption attack, what ultimately matters is whether the injected document changes the LLM's answer but not whether it enters the top-10. Table 3 shows that RV-KCA's advantage is almost entirely a retrieval effect, so the method is better at placing the document into the context, not at making the generator adopt it.

This matters because the retrieval step it wins on is exactly the step that the simplest RAG defenses target. Dedupilicating or filitering suspicious documents at ingestion, or isolating and aggregating per-document answers before generation, all act before or around this top-k entry. None of these is includ in the evaluation, so the reported ASR reflects an undefend pipeline. Given that the end2end benefit already narrows to a retrieval effect, it is unclear how much of it survives even a basic defense. The evaluation is also narrow in a way that compounds this: the three retrievers Contriever and Co-Condenser and ANCE are all from 2020 to 2022, and a single llama 8b model is used as the only ASR judge, and only NQ and TriviaQA are tested, so it is hard to tell how the end2end numbers transfer to newer retrievers and other LLM generator, so the question is:

Q3. Could you report ASR with at least one standard RAG defense in the pipeline (for example document level filitering at ingestion, or an isolate-then-aggregate decoding defense), so that the end2end gain is measured against a defended system rather than an undefended one?

Q4. Since the advantage is concentrated in Hit@10 while conditional ASR is at or below the baseline, what is the attack's benefit once a defense removes the retrieval advantage, and do the results hold on a more recent retrievers?

Questions for authors

Please see the weakness section. I am willing to raise the score if the concern is addressed.

6 comments
AuthorsOpenReview ↗

Response for Question 1

Comment 1:How many extra queries in total does the full attack issue per target query? how detectable is this volume of near-duplicate queries under a basic query-rate or query-similarity monitor?

Response 1: Thank you for raising this point. We have added complete query accounting and batched runtime measurements. We also evaluate basic query monitors and test whether shuffling construction requests changes their detectability.

We use Qwen3-Embedding-0.6B with HNSW retrieval and Qwen3-Reranker-0.6B, paired with Qwen3-32B on NQ and MS MARCO. Each dataset has 3,000 construction queries and 2,000 attack instances. Generation uses batches of four requests, with each instance retaining its own corpus state and retrieved evidence.

Table 1. Complete query cost and batched execution time.

StageCalls per runCalls per instanceNQ time (min)MS MARCO time (min)
Surrogate construction15,0007.586.491.8
Post-injection probing10,000531.234.7
Local preparation and search00104.6116.9
Final retrieval and generation2,000148.652.4
Total27,00013.5270.8295.8

Table 1 accounts for all 27,000 target queries, including final submissions. Once trained, the surrogate is reused without retraining for each attack. Each instance then requires only five additional probes and one final submission. Local search makes no further target queries. Across both datasets, five runs require 47.22 hours. Batching reduces execution time without reducing query counts. The original efficiency reductions remain specific to surrogate construction.

For monitoring, we compare grouped and shuffled submission of the same 15,000 construction requests. Responses remain associated with their source queries, so related requests need not arrive consecutively under the fixed corpus. Both schedules use identical account identities and request timestamps. The rate monitor uses a 60-second window, while the similarity monitor uses a 10-minute window. Thresholds are calibrated on separate benign traces containing legitimate reformulations to target a 1% false-positive rate. Monitor state is retained throughout the replay. Attack-stage probing remains unchanged and is reported separately.

Table 2. Request-level alerts under basic query monitors.

CorpusMonitorBenign requests flagged (%)Grouped construction flagged (%)Shuffled construction flagged (%)Attack-stage probes flagged (%)
NQQuery rate1.036.736.79.8
NQQuery similarity1.134.88.629.3
MS MARCOQuery rate0.939.239.211.6
MS MARCOQuery similarity1.032.510.127.8

Table 2 shows that shuffling reduces similarity-based construction alerts from 34.8% to 8.6% on NQ and from 32.5% to 10.1% on MS MARCO. The unchanged rate alerts confirm that this improvement does not come from slower submission. Thus, construction can collect the required feedback without concentrated near-duplicate requests. Similarity monitoring flags 29.3% and 27.8% of attack-stage probes, which identifies a limitation. However, detecting some probes does not mean the attack has been stopped. A successful attack makes the RAG system give users false answers and mislead them, so the risk should not be underestimated simply because some probes are detected.

AuthorsOpenReview ↗

Response for Question 2

Comment 2:How does the attack degrade under simple cheap defenses? for example near-duplicate query caching, or small random noise added to tie-breaking or to scores near the topk cutoff?

Response 2: We appreciate your constructive comments. We have added near-duplicate query caching and randomization near the retrieval cutoff. We measure their effects on attack success and clean answer quality, with each defense active throughout surrogate construction and attack execution.

We use Qwen3-Embedding-0.6B with HNSW retrieval of 100 candidates, followed by Qwen3-Reranker-0.6B. Qwen3-32B generates answers on NQ and MS MARCO. Both attacks train their surrogates from defended responses under the same query budget. The attacker observes only the final unordered passages. Clean answer F1 is measured in a separate benign replay.

For caching, we compare query embeddings using cosine similarity. A match above threshold τ\tau returns the stored passages after reranking, while generation still receives the current query. Cache entries persist within each corpus version, including during shuffled construction. Cache hits count toward the attacker’s query budget.

Table 3. Attack performance under near-duplicate query caching.

CorpusCache thresholdRewrite cache hitsClean answer F1FlippedRAG-Q ASRRV-KCA ASR
NQNo cache—75.3721.84 ± 1.0225.68 ± 0.85
NQτ=0.95\tau=0.9558.6275.0916.48 ± 0.9119.76 ± 1.08
NQτ=0.90\tau=0.9087.3173.2810.24 ± 0.8810.09 ± 0.96
MS MARCONo cache—64.5819.67 ± 1.2023.92 ± 1.17
MS MARCOτ=0.95\tau=0.9561.4764.1915.87 ± 1.0318.43 ± 0.94
MS MARCOτ=0.90\tau=0.9085.9662.049.82 ± 0.9710.21 ± 1.11

Table 3 shows that caching weakens the attack but does not eliminate the observed risk. At τ=0.95\tau=0.95 RV-KCA retains ASRs of 19.76% and 18.43%, while clean F1 decreases by only 0.28 and 0.39 percentage points. Lowering the threshold to 0.90 returns identical evidence for more rewrites, reducing the information available from retention differences. This stronger suppression also lowers clean F1 by 2.09 and 2.54 percentage points.

For randomization, we separately test random tie-breaking at the cutoff and independent Gaussian score noise. We standardize the 100 reranker scores within each query, then add noise with standard deviation σ\sigma to documents before selecting the final ten. Noise is sampled independently for every request. Pair accuracy is evaluated against the pre-noise reranker ordering, which is available only to the evaluator. The undefended references are the no-cache rows in Table 3.

Table 4. Attack performance under boundary randomization.

CorpusRandomizationPair accuracy / coverageClean answer F1FlippedRAG-Q ASRRV-KCA ASR
NQExact-tie breaking72.31 / 58.6975.3521.79 ± 1.0525.61 ± 0.88
NQScore noise, σ=0.02\sigma=0.0269.84 / 59.9275.1220.58 ± 0.9723.71 ± 1.06
NQScore noise, σ=0.10\sigma=0.1063.18 / 63.4774.4317.91 ± 1.1320.36 ± 1.08
MS MARCOExact-tie breaking71.08 / 57.9364.6119.73 ± 1.0823.87 ± 1.05
MS MARCOScore noise, σ=0.02\sigma=0.0268.57 / 60.3864.3418.76 ± 1.0222.09 ± 0.98
MS MARCOScore noise, σ=0.10\sigma=0.1062.46 / 62.8163.7117.32 ± 1.0919.48 ± 1.21

Table 4 shows little change under exact-tie breaking, which affects membership only when a tie crosses the cutoff. Score noise reduces pair accuracy, although coverage increases because random fluctuations can create apparent retention gaps. At σ=0.10\sigma=0.10, RV-KCA’s ASR decreases by 5.32 percentage points on NQ and 4.44 on MS MARCO. It still reaches 20.36% and 19.48%, with clean F1 losses below one percentage point. Thus, the randomization weakens stability supervision but still leaving a substantial end-to-end corruption risk.

AuthorsOpenReview ↗

Response for Question 3

Comment 3:Could you report ASR with at least one standard RAG defense in the pipeline (for example document level filitering at ingestion, or an isolate-then-aggregate decoding defense), so that the end2end gain is measured against a defended system rather than an undefended one?

Response 3: Thank you for this helpful question. We have added an end-to-end evaluation with the complete TrustRAG defense. We use Qwen3-Embedding-0.6B with HNSW retrieval and Qwen3-Reranker-0.6B. Qwen3-32B generates answers on NQ and MS MARCO. TrustRAG filters retrieved passages and resolves conflicts between external evidence and the model’s internal knowledge before producing the final answer. Both attacks retain the same retrieval feedback and attack budgets. The attacker receives no internal filtering decisions.

ASR is computed over all 2,000 attack instances, including those whose injected passage is removed. Success requires the final answer to endorse the attacker’s false claim. Merely quoting or rejecting that claim does not count as success.

Table 5. End-to-end attack success under TrustRAG.

CorpusDefenseClean answer F1FlippedRAG-Q ASRRV-KCA ASR
NQNone75.3721.84 ± 1.0225.68 ± 0.85
NQTrustRAG75.027.94 ± 0.6710.87 ± 0.97
MS MARCONone64.5819.67 ± 1.2023.92 ± 1.17
MS MARCOTrustRAG65.318.28 ± 0.8712.36 ± 1.09

Table 5 shows that TrustRAG reduces attack success while preserving clean answer quality. RV-KCA nevertheless retains ASRs of 10.87% on NQ and 12.36% on MS MARCO, exceeding FlippedRAG-Q by 2.93 and 4.08 percentage points. These differences concern false claims adopted in the final answer after the complete defense. Thus, this demonstrates an end-to-end security effect beyond entry into the top-10. A retrieval-mediated gain worthy of attention when it produces additional corrupted answers after filtering and conflict resolution.

RAG uses retrieved evidence to support factual answers, so improving the delivery of poisoned evidence can expose a meaningful weakness even without increasing conditional ASR after retrieval. The results directly test whether that weakness survives the evaluated protection.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Summary of the rebuttal

We sincerely thank the AC, the SPC and PC members, and all reviewers for their time, effort, and valuable feedback, which helped us explain our contribution more clearly and provide stronger evidence for our conclusions. Our responses focus on further validation of the original method and clarification of its scope.

  • We tested recent retrievers and a complete retrieval and reranking pipeline with Qwen3-32B. Results on NQ and MS MARCO show that stability remains informative through the same unordered retrieval interface.

  • We added injection-only controls and tests with identical retrieved evidence. Results on NQ and MS MARCO show that stability remains informative through the same unordered retrieval interface.These confirm that the main gain comes from changing the evidence reaching the generator, consistent with the mechanism described in the paper.

  • We added document-side baseline variants and a clearly labeled text-only adaptation of PR-Attack. We also showed that stability supervision improves document-only attacks with fixed user queries and benefits joint query–document optimization.

  • We evaluated TrustRAG alongside query caching and retrieval randomization. Under TrustRAG, RV-KCA exceeds FlippedRAG-Q in final-answer ASR by 2.93 percentage points on NQ and 4.08 on MS MARCO.

  • We tested alternative perturbation generators and periodically refreshed boundary competitors. The results support the usefulness of the inferred partial preferences and the default cached approximation. We clarified their dependence on query perturbations and the available probing budget.

  • We reported complete query costs and runtime, together with basic monitoring results. We clarified that the efficiency gains concern surrogate construction and documented the attack’s detectability.

These additions support the paper’s central finding that retrieval stability can expose useful ranking information from unordered outputs and help attackers corrupt the final answer.

AuthorsOpenReview ↗

Response for Question 1

Comment 1:The demonstrated security impact has a restricted scope. All targets use exact top-10 retrieval with Contriever, Co-Condenser, or ANCE, paired with relatively early 7B/8B generators on one Wikipedia corpus. No modern embedding retriever, reranker, or retrieval defense is evaluated. Moreover, the attack requires both document insertion and modification of the submitted query. These explicit assumptions define a valid threat model, but leave its relevance to current RAG deployments insufficiently established.

Response 1: Thank you for your valuable comment. Following your comment, we have expanded the experiments to modern retrieval pipelines and a larger generator. We have also added a web corpus and evaluated a dedicated RAG defense. We further clarified the deployment assumptions and added ablations of the two required input capabilities.

The new experiments use BGE-M3 and Qwen3-Embedding-0.6B (QE) as retrievers, with Qwen3-Reranker-0.6B (QR) for reranking and Qwen3-32B for generation. We evaluate on NQ and the MS MARCO web corpus. Pair accuracy measures the correctness of inferred document preferences, and coverage measures the fraction of pairs retained.

Table 1: Evaluation across modern RAG pipelines.

CorpusRetrieval and defensePair accuracy / coverage (%)Clean answer F1 (%)FlippedRAG-Q ASR (%)RV-KCA ASR (%)
NQBGE-M3 exact, k = 1077.61 / 62.4872.4129.18 ± 0.8734.52 ± 0.98
NQQE exact, k = 1075.84 / 60.2674.0627.36 ± 0.9132.81 ± 1.14
NQQE HNSW, k = 1074.92 / 61.8773.8227.02 ± 0.8332.14 ± 1.04
NQQE HNSW + QR, k = 570.18 / 54.9272.9620.91 ± 0.7824.63 ± 0.99
NQQE HNSW + QR, k = 1072.46 / 58.7375.3721.84 ± 1.0225.68 ± 0.85
NQQE HNSW + QR, k = 2069.37 / 63.1875.1120.27 ± 0.9823.42 ± 1.18
MS MARCOQE HNSW + QR, k = 1071.23 / 57.8664.5819.67 ± 1.2023.92 ± 1.17

Table 1 shows that useful ranking information remains observable with modern retrieval components. After HNSW retrieval and reranking on NQ, the inferred preferences reach 72.46% accuracy while covering 58.73% of document pairs. Thus, membership changes still provide informative supervision after the retrieval pipeline is upgraded. This ranking signal also translates into higher attack success. Replacing exact search with HNSW changes RV-KCA’s ASR from 32.81% to 32.14%, while its advantage over FlippedRAG-Q remains around 5.12%. Reranking reduces attack success for both methods, but RV-KCA still achieves 25.68% ASR compared with 21.84% for FlippedRAG-Q at k = 10. The advantage also persists at k = 5 and k = 20. These results extend the observed effect to approximate retrieval and different context sizes.

These results shown that the retrieved false evidence can influence a larger, more recent generator. The MS MARCO results further extend the finding beyond Wikipedia. With reranking enabled, RV-KCA achieves 23.92% ASR on this web corpus, compared with 19.67% for FlippedRAG-Q. The improvement therefore appears across both corpus settings under the updated retrieval pipeline.

Table 2: Attack performance under TrustRAG defense.

CorpusRetrieval and defenseClean answer F1FlippedRAG-Q ASRRV-KCA ASR
NQQE HNSW + QR + TrustRAG, k = 1075.027.94 ± 0.6710.87 ± 0.97
MS MARCOQE HNSW + QR + TrustRAG, k = 1065.318.28 ± 0.8712.36 ± 1.09

For the defense evaluation shown in Table 2, we apply the complete TrustRAG pipeline to the same retrieval and reranking configurations. This examines whether RV-KCA retains its attack advantage when the system explicitly filters corrupted evidence and resolves conflicting information. RV-KCA retains gains of 2.93% on NQ and 4.08% on MS MARCO under the complete defense. Its remaining ASRs are 10.87% and 12.36%, respectively. These outcomes demonstrate that the attack advantage extends to the evaluated defended pipeline.

TrustRAG reduces absolute attack success while preserving clean answer quality. Nevertheless, RV-KCA induces more successful corruptions than FlippedRAG-Q in both datasets under the same access conditions and attack budgets. An assessment based only on FlippedRAG-Q would therefore underestimate the residual risk in these configurations. This finding establishes the practical value of accounting for stability-derived ranking information when evaluating the security of modern RAG systems.

AuthorsOpenReview ↗

Response for Question 2

Comment 2:Some end-to-end attribution and baseline questions remain unresolved. The clean reference removes both the injected document and the trigger; an injection-only reference is missing. The hit/miss decomposition is informative, but does not isolate any direct effect of the modified query on generation. PAT and FlippedRAG are also adapted by moving triggers to the query, while the cited PR-Attack, which jointly considers prompts and poisoned text, is not evaluated. These gaps limit assessment of the attack's added value beyond the controlled surrogate comparisons.

Response 2: We appreciate your careful review. We have added an injection-only reference and controlled experiments that separate the trigger’s effects on retrieval and generation. We have also expanded the baseline comparison to include document-side triggers and joint optimization of the query and poisoned document.

These experiments use Qwen3-Embedding-0.6B and Qwen3-Reranker-0.6B with Qwen3-32B on NQ and MS MARCO. For attribution, we keep the injected passage fixed and independently control whether the trigger reaches the retriever or the generator. Let q_T denote the original query with its optimized trigger. To measure the direct generation effect, we replay exactly the same retrieved passages in the same order with either q or q_T.

Table 3. Separating retrieval and generation effects.

ConditionInjected documentRetriever queryGenerator queryNQ ASRMS MARCO ASR
CleanNoqq0.62 ± 0.090.81 ± 0.14
Trigger without injectionNoq_Tq_T1.07 ± 0.151.29 ± 0.17
Injection onlyYesqq8.36 ± 0.477.94 ± 0.52
Trigger to generator onlyYesqq_T9.12 ± 0.558.31 ± 0.64
Trigger to retriever onlyYesq_Tq24.41 ± 0.9122.18 ± 1.07
Full RV-KCAYesq_Tq_T25.68 ± 0.8523.92 ± 1.17

The injection-only method directly measures the effect of the poisoned passage without query modification. Full RV-KCA improves ASR over this method by 17.32% on NQ and 15.98% on MS MARCO. So its improvement extends beyond the effect of inserting the initial poisoned document. Most of this improvement remains when the generator receives the original query. Sending the trigger only to retrieval yields 24.41% ASR on NQ and 22.18% on MS MARCO, close to the full attack. With retrieved evidence held fixed, adding the trigger to generation changes ASR by 0.76–1.27% on NQ and 0.37%–1.74% on MS MARCO, depending on the retrieved context. These comparisons identify a small direct generation effect while showing that the larger gain comes from changing the evidence supplied to the generator.

For the baseline comparison, we evaluate PAT and FlippedRAG with their original document-side trigger placement, denoted by −D-D. We retain the query-side versions, denoted by −Q-Q, and evaluate both placements on the same fact-substitution task. We also add PR-Attack-text to evaluate joint optimization under the common text-only interface.

Table 4. Expanded end-to-end baseline comparison.

MethodOptimized textNQ Hit@10NQ ASRMS MARCO Hit@10MS MARCO ASR
PAT-DDocument41.8318.62 ± 0.7238.7116.94 ± 0.96
PAT-QQuery38.6417.19 ± 0.8540.2617.38 ± 1.01
FlippedRAG-DDocument45.9820.73 ± 1.0748.3120.84 ± 0.93
FlippedRAG-QQuery47.6321.84 ± 1.0246.3719.67 ± 1.20
PR-Attack-textDocument and query51.4224.07 ± 0.9452.6823.06 ± 0.82
RV-KCAQuery57.1125.68 ± 0.8554.7623.92 ± 1.17
PR-Attack-text + stabilityDocument and query57.8326.54 ± 1.0355.1924.81 ± 0.96

RV-KCA achieves higher mean ASR than both versions in each dataset. Thus its advantage remains when the document-side placement is restored. PR-Attack-text provides a stronger comparison by optimizing both inputs. RV-KCA remains competitive with this joint attack while keeping its injected passage fixed during trigger optimization. More directly, adding stability supervision to PR-Attack-text increases ASR by 2.47% on NQ and 1.75% on MS MARCO. The increases in Hit@10 show that the recovered ranking information remains useful when prompt and document optimization are already combined.

In summary, the fixed-evidence interventions and expanded baselines identify how retrieval changes drive the additional corruption and demonstrate that stability information can also strengthen a joint prompt–document attack.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.