acceptodds
Under review as a conference paper at ICLR 2027

GE: Grounded Exploitation–Exploration Extraction from Black-Box RAG Systems

Abstract

Retrieval-Augmented Generation (RAG) systems combine Large Language Models (LLMs) with external knowledge repositories, creating a new attack surface in which a proprietary corpus can be systematically extracted through conversational interaction. Explicit extraction attacks that directly request verbatim content are readily detected by lexical guardrails and refusal mechanisms, motivating stealthier semantic extraction through benign queries. Yet existing implicit probing methods remain vulnerable to semantic drift and localized exploration, limiting their ability to systematically cover the underlying corpus. In this work, we show that grounded extraction provides an effective design principle for addressing these limitations: generating candidate attack queries directly from returned responses mitigates out-of-domain semantic drift, while selecting attack queries via history-relative maximin distance optimization prevents extraction from becoming trapped within localized concept clusters. We further identify a complementary leakage channel arising from the discrepancy between retrieval and generation, which we term Echo Knowledge Overflow (EKO). Because production RAG systems often retrieve more passages than are needed to answer a query, portions of the retrieved context may remain unused in the response. However, this latent context can be surfaced as semantically adjacent follow-up cues through benign continuation-style instructions. Building on these observations, we introduce Grounded Exploitation–Exploration Extraction (GE), a black-box knowledge extraction attack that operates through benign natural-language queries. GE implements a dual-track strategy. Its exploration track derives progressive queries from facts confirmed in responses, while its exploitation track harvests EKO-derived cues using inductive continuation suffixes. A history-relative maximin distance algorithm then selects the next query from the response-grounded candidate pool to balance local exploitation with broader corpus exploration under a constrained query budget. Across four benchmark corpora and twelve victim LLMs, GE achieves macro-average Retrieval Coverage Rates of 0.847 on local models and 0.846 on enterprise endpoints within a 1,000-query budget, outperforming evaluated baselines by 15.1 and 21.3 percentage points, respectively, while sustaining a 96.8–97.6% Attack Success Rate against input-output guardrails. Beyond exposing retrieved knowledge, the extracted information enables the construction of Shadow RAG surrogates that replicate victim services with up to an 84.7% valid answer rate and 0.812 semantic answer similarity, demonstrating the practical feasibility of functional service replication.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-Augmented Generation, Knowledge Extraction, Data Privacy, AI Safety, Echo Knowledge Overflow

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 4 reviews, above 68% of the others.
Reviewer_PDoUOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

GE³ is a black-box attack that extracts a RAG system’s private corpus with benign queries, follow-up question prompts, and distance-based query selection. On four corpora and twelve LLMs, it covers more of the corpus than three baselines and gives enough data to copy the service.

Critical strengths

The experimental part is quite solid with many models and proper ablation, including six commercial models and three pipeline defenses, with the 2x2 ablation cleanly separating the contribution of the dual track from that of maximin selection.

Critical weaknesses

There is no limitation in the main body. One of the main contribution of the paper is EKO, while we cannot determine either the unused retrieved content was actually leaked, or the model is simply making associations based on the same topic. Besides, we cannot distinguish extracted content and retrieved content. The coverage of GE^3 do not show clear advantage as baseline (Dual-Rand and Ans-MMD).

Questions for authors

Look at critical Weakness section

3 comments
AuthorsOpenReview ↗

Response to Reviewer PDoU [1/2]

We sincerely appreciate your constructive comments and valuable feedback. Below, we provide point-by-point responses supported by additional analyses and experiments. We would be happy to provide further clarification or supporting evidence if needed.

Critical weakness, point 1. Limitations in the main body

Thank you for highlighting the importance of explicitly discussing limitations. We have substantially revised Section 5\color{blue}{\text{Section 5}} to include a dedicated limitation discussion in the main manuscript.

Specifically, we distinguish evaluator-visible retrieval coverage from actual information disclosure, motivating future methods for directly measuring recovered knowledge from public responses. We also discuss the distinction between semantic novelty and knowledge discovery, as well as opportunities for adaptive coordination between answer-grounded exploration and EKO-driven exploitation. These directions extend the current framework toward more knowledge-aware and adaptive extraction.


Critical weakness, point 2. EKO leakage or topic association?

We appreciate the concern that EKO continuation cues may simply reflect topical associations rather than information from retrieved passages. To address this, we tested eight questions per corpus from Healthcare, TREC-COVID, SciDocs, and NFCorpus, using ten retrieved passages, BGE-M3 and Qwen3-Embedding-8B, with complete context, single-passage deletion, length-matched same-corpus replacement, and empty context (Appendix B.2\color{blue}{\text{Appendix B.2}}).

Rather than relying on semantic similarity, we identify source-specific five-word phrases shared by a passage and continuation but absent from the query, primary answer, and other passages. We use five words to help distinguish source-specific expressions from incidental short matches while retaining matches in brief continuation questions. Deletion reduces these signals in 6/8, 5/8, and 6/8 literature-corpus questions using BGE-M3, and 7/8, 6/8, and 7/8 using Qwen (Table B.1\color{blue}{\text{Table B.1}}).

A further control holds the complete-context primary answer fixed, ruling out changes in that answer as the sole explanation. For the 24 literature-corpus questions with BGE-M3, the same answer excludes 53 of 240 question–passage positions sharing a source-specific five-word phrase with it, leaving 187 per mode, fixed across interventions. Source matches fall from 48/187 to 0/187 after deletion in ordinary generation, and from 44/187 to 3/187 with the primary answer fixed; replacement introduces 35/187 and 45/187 new-source matches (Appendix B.2\color{blue}{\text{Appendix B.2}}, “Primary-answer overlap”). These findings demonstrate that at least some EKO cues depend on retrieved information beyond the primary answer, rather than merely reflecting topical associations.

AuthorsOpenReview ↗

Response to Reviewer PDoU [2/2]

Critical weakness, point 3. Extracted content versus retrieved content

Retrieving a corpus record does not necessarily imply that its contents have been disclosed. Retrieval coverage (CR) measures the fraction of corpus records retrieved using evaluator-visible identifiers, whereas the attacker has access only to generated responses. The revised Section 4.1\color{blue}{\text{Section 4.1}} defines CR as a retrieval metric rather than a measure of disclosed information.

To assess the utility of information actually available to the attacker, we added a 100-question Healthcare evaluation (Section 4.3\color{blue}{\text{Section 4.3}}). Shadow RAG uses public answers from 1,000 target requests, without access to original records or internal retrieved passages. Both conditions use the same Llama-3.1-8B generator. On questions targeting extracted facts verified against original records, necessary-fact completeness increases from 18/100 (closed-book) to 92/100 (Shadow RAG). Here, necessary facts are the key pieces of information that an answer must include, established from the source record before evaluation. Completeness requires all essential source-supported facts to be correctly conveyed. Since the generator remains identical, these results demonstrate that the recovered information provides substantial factual utility beyond the generator's unaided responses, complementing the source-dependence evidence from the controlled interventions.


Critical weakness, point 4. Coverage relative to Dual-Rand and Ans-MMD

Our component analysis (Table E.2\color{blue}{\text{Table E.2}}) reveals a coverage–fidelity–acceptance trade-off and clarifies the distinct contributions of EKO and maximin selection.

GE³ achieves average CR of 84.71%, higher than Dual-Rand (77.99%) but lower than Ans-MMD (87.65%). Maximin selection improves coverage over Dual-Rand in 23/24 configurations, while introducing EKO to Ans-MMD improves SF and ASR in all 24 configurations, despite reducing coverage in 18.

Specifically, adding EKO increases average SF from 0.6679 to 0.7365 and ASR from 90.15% to 97.57%, while decreasing CR from 87.65% to 84.71%. These results show that EKO improves extraction fidelity and answer acceptance, whereas maximin promotes broader retrieval coverage within the dual-track framework.

We have clarified this distinction in the revised Section 4.4\color{blue}{\text{Section 4.4}}, emphasizing that GE³ achieves a competitive balance across the three objectives rather than maximizing coverage alone.

We sincerely appreciate your time and effort in reviewing our paper. We have provided detailed responses to your comments, along with additional experiments and revisions to address your concerns.

As the discussion period is drawing to a close, we would be very grateful if you could take a moment to review our responses when convenient. If you have any remaining concerns or would like further clarification, we would be more than happy to address them.

Thank you again for your valuable feedback and consideration.

Reviewer_n4i5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GE3, a black-box knowledge extraction attack against RAG systems through benign queries. It combines two query generation strategies: answer-grounded exploration, which generates follow-up queries from previously obtained responses. And the EKO-driven exploitation, which uses continuation prompts to elicit additional query cues potentially derived from retrieved but unexpressed context. It also adopts a history-relative maximin distance strategy to select diverse queries to improve corpus coverage under a limited query budget. Experiments show that GE3 outperforms the baselines in retrieval coverage and guardrail bypass rates. The authors further demonstrate that the collected corpus can be used to construct Shadow RAG systems with similar answering behavior.

Critical strengths
  1. This paper investigates a relevant threat to RAG systems, showing how benign interactions may facilitate knowledge extraction and unauthorized service replication without requiring internal access.

  2. GE^3 combines response-grounded query generation, continuation-based cue extraction, and diversity-aware query selection into a coherent black-box framework. It is relatively simple to implement and does not require access to victim model parameters.

  3. The reported improvements over IKEA and RAG-Crawler are substantial.

Critical weaknesses
  1. The main components of GE^3 substantially overlap with existing techniques. Response-guided query generation and iterative exploration have been investigated in prior work, including RAG-Thief/CopyBreakRAG and IKEA, while diversity-aware exploration and corpus coverage optimization have been studied in RAG-Crawler and GeoEx.

  2. Section 2.2 categorizes existing defenses into three types: Input/Output Perimeters, Runtime Monitors, and Pipeline Interventions. However, Section 4.5 evaluates only three pipeline-level defenses (Query Rewriting, DPRAG, and SAGE), in addition to the basic input/output guardrails used throughout the experiments. The authors do not evaluate runtime monitoring defenses, such as CanaryRAG, ControlNet, or RAGFort, despite discussing them in the related work. Since GE^3 relies on repeated, adaptive querying, its effectiveness against defenses that monitor interaction patterns or detect suspicious extraction behavior remains unclear.

  3. The paper claims that EKO enables the extraction of follow-up cues from retrieved context that remain unexpressed in the primary answer. However, the evidence does not sufficiently support this mechanism. First, the "unused" chunks are identified based on their low semantic similarity to the primary answer, and the authors subsequently demonstrate that continuation spans exhibit higher similarity to these same chunks. This evaluation introduces a potential selection bias, as the primary answer is already expected to have low similarity to the selected subset. Second, the generated follow-up questions may simply reflect the LLM's knowledge, rather than information derived from unused retrieved passages.

Questions for authors
  • Would a controlled experiment that removes or replaces the unused retrieved chunks, with everything else held fixed, show that EKO exploits latent retrieved context rather than ordinary follow-up question generation?
5 comments
AuthorsOpenReview ↗

Response to Reviewer n4i5 [1/3]

We sincerely appreciate your constructive comments and valuable feedback. Below, we provide point-by-point responses supported by additional analyses and experiments. We would be happy to provide further clarification or supporting evidence if needed.

Weakness 1. Novelty Relative to Existing Extraction Methods

Our contribution lies in identifying and exploiting an additional source of retrieval-grounded information, and integrating it with answer-grounded exploration through history-relative query selection. The key distinction is that GE³ extends adaptive query generation beyond information explicitly expressed in previous answers, leveraging victim-generated continuation cues that can reflect otherwise unexpressed retrieved context.

Compared with existing methods, RAG-Thief and GeoEx employ explicit context-reproduction instructions, making them susceptible to input and lexical output defenses. Among implicit extraction methods, IKEA generates follow-up queries through response-derived anchor mutation, while RAG-Crawler traverses entity-relation graphs constructed from previous answers. Although these approaches employ different exploration strategies, their subsequent queries primarily rely on information already expressed in observable responses.

Black-box attacks commonly use returned feedback for local computation, memory updates and iterative querying. GE³ processes that feedback through three complementary mechanisms. First, EKO elicits victim-generated follow-up cues that can reflect retrieved information absent from the primary answer, extending candidate generation beyond answer-visible facts. Second, answer-grounded exploration and EKO-driven exploitation provide distinct sources of grounded candidates without requiring predefined anchors or graph connectivity. Third, history-relative maximin selection promotes semantic diversity within the response-grounded candidate pool, reducing repetitive exploration without introducing unconstrained semantic drift. These choices form a distinct extraction mechanism that operates through ordinary content questions and a benign continuation request.

These distinctions are supported by additional experiments. In the common 187-position BGE-M3 sensitivity set, source matches decrease from 44/187 to 3/187 after passage deletion with the primary answer fixed, while replacement introduces 45/187 new-source matches (Appendix B.2\color{blue}{\text{Appendix B.2}}). Factorial ablations show that adding EKO under maximin selection increases average SF from 0.6679 to 0.7365 and ASR from 90.15% to 97.57%, with SF and ASR improving in all 24 configurations. Maximin increases coverage from 77.99% (Dual-Rand) to 84.71% (GE³), while GE³ has higher mean harmonic utility than Ans-MMD (0.8411 versus 0.7988), despite lower coverage (84.71% versus 87.65%) (Table E.2\color{blue}{\text{Table E.2}}; Section 4.4\color{blue}{\text{Section 4.4}}).

AuthorsOpenReview ↗

Response to Reviewer n4i5 [2/3]

Weakness 2. Evaluation against additional runtime defenses

We appreciate this suggestion and have extended our evaluation to include ControlNET, CanaryRAG, and RAGFort using the Healthcare corpus, covering activation-based detection, canary-based monitoring, and generation-stage intervention, respectively (Section 4.5\color{blue}{\text{Section 4.5}}).

ControlNET: On the first 100 requests per attack, GE³ achieves a detection rate of 17%, compared with 40% for IKEA, 75% for RAG-Crawler, and 85% for RAG-Thief (Figure F.1\color{blue}{\text{Figure F.1}}). GE³ is less frequently flagged by the evaluated detector. The shared-prefix, single-trajectory setup is described in Appendix F.4\color{blue}{\text{Appendix F.4}}.

CanaryRAG: Among 200 GE³ requests, 11 (5.5%) are blocked, while coverage decreases from 63.0% to 59.4% and SF from 0.7207 to 0.7148 (Table F.3\color{blue}{\text{Table F.3}}). CanaryRAG blocks 13/200 IKEA requests (6.5%), 20/200 RAG-Crawler requests (10.0%), and 57/200 RAG-Thief requests (28.5%), including both target-canary and oracle checks. On a separate workload of 64 ordinary questions, it produces six false blocks (9.38%), illustrating the trade-off between attack suppression and legitimate-query acceptance (Appendix F.3\color{blue}{\text{Appendix F.3}}).

RAGFort: With its cascade and critic enabled, GE³'s coverage decreases substantially from 63.0% to 34.3%. Under the same defense, IKEA, RAG-Crawler, and RAG-Thief achieve coverage of 33.3%, 28.0%, and 9.1%, respectively (Table 4\color{blue}{\text{Table 4}}). Although RAGFort substantially reduces retrieval coverage, GE³ retains the highest coverage among the evaluated attacks.

These experiments demonstrate that runtime defenses can meaningfully reduce extraction effectiveness, particularly RAGFort. Nevertheless, GE³ retains measurable coverage under the evaluated defenses. We have incorporated these findings into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer n4i5 [3/3]

Weakness 3, Question 1. Evidence for the EKO Mechanism

We appreciate the reviewer's two concerns and address them separately.

First, defining latent context based on low similarity to the primary answer does not, by itself, establish that continuation cues originate from those passages. We therefore conducted passage-deletion and length-matched same-corpus replacement experiments on eight questions each from Healthcare, TREC-COVID, SciDocs, and NFCorpus, using ten retrieved passages and both BGE-M3 and Qwen3-Embedding-8B (Appendix B.2\color{blue}{\text{Appendix B.2}}). Complete and empty contexts provide the controls.

Rather than relying on semantic similarity, we identify exact five-word phrases shared by a source and continuation but absent from the query, primary answer, and other passages. Five words are chosen to reduce incidental short matches while retaining matches in brief follow-up questions. Deletion reduces source-specific signals in 6/8, 5/8, and 6/8 literature-corpus questions with BGE-M3, and 7/8, 6/8, and 7/8 with Qwen (Table B.1\color{blue}{\text{Table B.1}}).

Holding the complete-context primary answer fixed further tests whether changes in that answer alone explain the changed cues. In the BGE-M3 sensitivity analysis on the 24 literature-corpus questions, that same answer excludes 53 of 240 question–passage positions sharing a source-specific five-word phrase with it, leaving 187 in both modes, fixed across interventions. Deletion reduces matches from 48/187 to 0/187 in ordinary generation and 44/187 to 3/187 with the primary answer fixed; replacement introduces 35/187 and 45/187 new-source matches, respectively (Appendix B.2\color{blue}{\text{Appendix B.2}}). These findings support source dependence beyond the original similarity-based grouping or changes in the primary answer alone.

Second, regarding whether the observed information merely reflects the LLM's pretrained knowledge, our passage-replacement experiments provide direct evidence that continuation cues depend on the supplied retrieved context: replacing a source introduces corresponding new-source phrases despite using the same underlying generator.

We conducted a 100-question Healthcare evaluation to isolate the factual contribution of extracted knowledge (Section 4.3\color{blue}{\text{Section 4.3}}). Both conditions use Llama-3.1-8B as the generator, and Shadow RAG relies exclusively on public answers collected from 1,000 target requests. On 100 questions targeting facts contained in GE³'s extracted corpus and verified against original records, providing the extracted corpus increases necessary-fact completeness from 18/100 (closed-book) to 92/100 (Shadow RAG). Here, necessary facts are the key pieces of information that an answer must include, established from the source record before evaluation. Completeness requires all essential source-supported facts to be correctly conveyed. Since the generator remains identical, these results demonstrate that the recovered information provides substantial factual utility beyond the generator's unaided responses, complementing the source-dependence evidence from the controlled interventions.

Reviewer_zo3eOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes GE3, a black-box knowledge extraction framework for RAG systems. It combines answer-grounded exploration, EKO-driven exploitation, and history-relative maximin query selection to improve extraction coverage under a limited query budget. Experiments across four datasets and twelve LLMs show substantial improvements over existing baselines, achieving approximately 0.85 retrieval coverage within 1,000 queries.

Critical strengths
  1. Important problem: The paper addresses practical knowledge leakage risks in black-box RAG systems.

  2. Interesting framework: Combining response-grounded candidate generation with maximin query selection is intuitive and technically lightweight.

  3. Extensive experiments: Evaluation covers multiple datasets, model families, ablations, and defense configurations.

  4. Promising results: GE3 consistently outperforms the evaluated baselines in the main experiments.

Critical weaknesses
  1. Insufficient evidence for EKO. Appendix B defines unused context based on its low similarity to the primary answer. This introduces potential selection bias. The observed similarity shift does not conclusively demonstrate that continuation cues originate from otherwise unused retrieved passages. Controlled experiments are needed to establish this mechanism.

  2. Ablation results raise questions. Ans-MMD achieves higher coverage (0.877) than full GE3 (0.847), while Dual-Rand also performs worse in coverage than Ans-Rand. This suggests that EKO may improve semantic fidelity and query acceptance rather than coverage itself. The authors should clarify this trade-off.

  3. Shadow RAG evaluation is incomplete. Semantic similarity to victim responses does not necessarily establish factual correctness or successful service replication. Independent ground-truth evaluation would strengthen this claim.

Questions for authors
  1. Can the authors measure actual information recovered in attacker-visible responses rather than evaluator-visible retrieval coverage?

  2. Can controlled experiments establish that EKO cues specifically originate from unused retrieved passages?

  3. Why does the full GE3 framework achieve lower coverage than Ans-MMD in the ablation study?

  4. How much of Shadow RAG's performance comes from extracted knowledge rather than the pretrained generator's existing capabilities?

4 comments
AuthorsOpenReview ↗

Response to Reviewer zo3e [1/2]

We greatly appreciate your thoughtful comments and valuable feedback. We provide detailed, point-by-point responses below, incorporating additional experiments and analyses to address your concerns. We would be happy to provide further clarification or supporting evidence if needed.

Weakness 1, Question 2. Controlled evidence for EKO

We appreciate the concern regarding potential selection bias in our original similarity-based analysis. Grouping passages according to their similarity to the primary answer does not, by itself, establish that continuation cues depend on those passages. We therefore conducted controlled interventions that directly manipulate the retrieved context (Appendix B.2\color{blue}{\text{Appendix B.2}}).

For eight questions per dataset across TREC-COVID, SciDocs, and NFCorpus, we compare complete-context, individual passage-deletion, length-matched replacement, and empty-context conditions using BGE-M3 and Qwen3-Embedding-8B. We identify source-specific information through five-word phrases appearing in a retrieved passage and continuation question, but absent from the original query, primary answer, and other supplied passages. We use five-word phrases to reduce incidental matches while retaining matches in brief follow-up questions. If EKO cues depend on the retrieved context, removing the corresponding passage should reduce the match, while replacing it should introduce signals associated with the replacement. The results demonstrate clear source dependence. Passage removal reduces source-specific continuation signals in 6/8, 5/8, and 6/8 questions across the three corpora using BGE-M3, and 7/8, 6/8, and 7/8 using Qwen (Table B.1\color{blue}{\text{Table B.1}}). Replacement similarly introduces cues associated with the newly supplied passages.

A further control holds the primary answer fixed while changing the passages used to generate the follow-up. This matters because it rules out changes in the primary answer as the sole explanation for the changed follow-up. In ordinary generation, the model generates both the primary answer and follow-up questions. In the fixed-primary-answer control, that complete-context primary answer is supplied unchanged. For BGE-M3 on the 24 questions, the same primary answer defines both modes' exclusion set: 53 of 240 question–passage positions share a source-specific five-word phrase with it, leaving 187, held fixed across interventions. Source matches fall from 48/187 to 0/187 after deletion in ordinary generation, and from 44/187 to 3/187 in the fixed-primary-answer control; replacement introduces 35/187 and 45/187 new-source matches, respectively (Appendix B.2\color{blue}{\text{Appendix B.2}}, “Primary-answer overlap”). These results provide further evidence that EKO continuation cues can depend on retrieved information beyond what is expressed in the primary answer, rather than merely reflecting changes in the primary answer itself.

AuthorsOpenReview ↗

Response to Reviewer zo3e [2/2]

Weakness 2, Question 3. Why Does GE³ Have Lower Coverage Than Ans-MMD?

EKO does not maximize retrieval coverage in isolation. Instead, its contribution lies in improving semantic fidelity and answer acceptance. Our component ablation analysis (Table E.2\color{blue}{\text{Table E.2}}) shows that introducing EKO under identical maximin selection reduces average CR from 87.65% to 84.71%, while increasing SF from 0.6679 to 0.7365 and ASR from 90.15% to 97.57%. SF and ASR improve across all 24 dataset–generator configurations. This establishes a measurable trade-off: answer-only exploration can achieve greater retrieval breadth, while EKO-derived cues produce more faithful and readily accepted responses. Such improvements do not necessarily translate into additional unique retrieved passages.

Importantly, maximin selection partially compensates for the coverage reduction, increasing CR from 77.99% (Dual-Rand) to 84.71% (GE³). The revised Section 4.4\color{blue}{\text{Section 4.4}} explicitly distinguishes EKO's contribution to fidelity and acceptance from maximin's contribution to coverage.


Weakness 3, Question 4. Shadow RAG Factual Utility Beyond Pretrained Knowledge

We agree that semantic similarity alone cannot establish factual correctness or successful service replication. We therefore added a 100-question Healthcare evaluation with independent source-based factual assessment (Section 4.3\color{blue}{\text{Section 4.3}}). Shadow RAG uses only public answers collected from 1,000 target requests. We keep Llama-3.1-8B as the generator throughout, so the comparison measures what the extracted text adds to the model's unaided answers.

On 100 questions targeting facts contained in GE³'s extracted corpus and verified against original records, providing the extracted corpus increases necessary-fact completeness from 18/100 (closed-book) to 92/100 (Shadow RAG). Here, necessary facts are the key pieces of information that an answer must include, established from the source record before evaluation. Completeness requires all essential source-supported facts to be correctly conveyed. Since the generator remains identical, these results demonstrate that the extracted knowledge provides substantial factual utility beyond the model's unaided responses.

We report this additional evaluation in the revised Section 4.3\color{blue}{\text{Section 4.3}}.


Question 1. Attacker-Visible Information Versus Retrieval Coverage

We agree that retrieval coverage does not necessarily indicate actual information disclosure. CR measures the fraction of corpus records retrieved during an attack using evaluator-visible retrieval identifiers, whereas a black-box attacker has access only to the generated responses. Thus, retrieving a record does not necessarily mean that its information has been exposed to the attacker.

To complement CR, we evaluate the factual utility of attacker-visible knowledge through the additional Healthcare experiment described in our response to Weakness 3. Shadow RAG is constructed exclusively from public answers collected during the attack, without access to internal retrieval results or the original corpus. Using the same generator, necessary-fact completeness increases from 18/100 (closed-book) to 92/100 (Shadow RAG) on source-verified questions targeting extracted facts. This demonstrates that the collected responses contain reusable factual information beyond the model's unaided answers. The revised Section 4.3\color{blue}{\text{Section 4.3}} reports the complementary factual evaluation.

Thank you for the detailed response and the substantial experimental additions.

The controlled passage-manipulation experiments in Appendix B.2 (particularly the fixed-primary-answer control) convincingly establish that EKO continuation cues depend on retrieved passages beyond the primary answer. Furthermore, the clarified discussion in Section 4.4 regarding the trade-off between retrieval coverage (driven by maximin) and semantic fidelity/acceptance rate (driven by EKO) fully resolves my question on the ablation results.

Regarding W3, Q1, and Q4, the new 100-question Healthcare evaluation (Section 4.3) showing an increase from 18/100 (closed-book) to 92/100 (Shadow RAG) effectively demonstrates that attacker-visible responses contain genuine factual utility beyond the base generator's parametric knowledge. As a minor note for the final revision: since these 100 questions specifically target facts within GE3's extracted corpus, this metric reflects factual fidelity conditioned on successful extraction rather than end-to-end factual coverage over the entire victim database. Briefly noting this distinction in Section 4.3 will make the scope of the evaluation completely transparent.

Overall, my primary concerns have been well addressed, and I appreciate the authors' effort in strengthening the paper.

We sincerely thank you for your thoughtful follow-up and encouraging assessment of our revisions. We are very pleased that the additional controlled experiments and clarifications have addressed your primary concerns.

We also appreciate your suggestion regarding the scope of the 100-question Healthcare evaluation. We agree that this evaluation measures factual completeness conditioned on successfully extracted knowledge, rather than end-to-end factual coverage of the entire victim corpus. We have explicitly clarified this distinction in Section 4.3 of the final revision.

Given that your primary concerns have been addressed, we would be very grateful if you would consider revisiting your overall rating in light of the additional experimental evidence and revisions. We sincerely appreciate your time and constructive feedback throughout the review process.

Reviewer_x4YSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies an important security problem in black-box RAG systems. It proposes GE³, which combines answer-grounded exploration, EKO-driven exploitation, and maximin query selection.

Critical strengths
  1. Intuitive framework. The two tracks provide different sources of follow-up questions, helping explore the knowledge base. Maximin encourages query diversity while keeping questions grounded in previous responses.

  2. Solid experiments. The paper evaluates four datasets and twelve victim LLMs. The results are promising, with additional ablation, defense, and Shadow RAG experiments.

Critical weaknesses
  1. The contribution of EKO is still unclear. In the ablation (Lines 419–431), answer-only methods achieve higher coverage: Ans-MMD (0.877) and Ans-Rand (0.862), compared with GE³ (0.847). GE³ performs better on the combined metric, but the harmonic mean treats CR, SF, and ASR equally. Why is this weighting appropriate?

  2. The balance between the two query tracks is unclear. Both tracks share one candidate pool, and maximin selects queries without considering their source (Lines 238–240). Could one track dominate the selection, leaving the other rarely used? Reporting the selection ratio and contribution of each track would help clarify whether both tracks are effectively utilized.

Questions for authors
  1. EKO mechanism: Does appending the continuation suffix change the retrieved passages? Could the authors compare retrieval results with and without the suffix?

  2. EKO contribution and track balance: Why are CR, SF, and ASR equally weighted in the harmonic mean? Would different weights change the conclusions? What proportion of the executed queries comes from each track?

  3. Shadow RAG evaluation: How does Shadow RAG compare with the same LLM without retrieval? How much improvement actually comes from the extracted knowledge rather than the pretrained knowledge?

  4. Minor Comments: Abstract (Lines 40–43): “84.7% valid answer rate” should be clarified as a relative response rate. Appendix E (Figure E.1): LLM-Rand/LLM-MMD and Ans-Rand/Ans-MMD are used inconsistently. Line 489: “Appendix A, and Appendix H” → “Appendices A and H.”

3 comments
AuthorsOpenReview ↗

Response to Reviewer x4YS [1/2]

We sincerely appreciate your constructive comments and valuable feedback. Below, we provide point-by-point responses to your concerns and questions, supported by additional analyses and experimental evidence. Please let us know if any points require further clarification or if additional evidence would be helpful.

Weakness 1(a). EKO Contribution and Coverage Trade-off

EKO improves extraction fidelity and answer acceptance rather than maximizing retrieval coverage alone. Unlike answer-grounded exploration, which generates follow-up questions from explicitly returned facts, EKO provides an additional source of candidates through the victim's retrieval-conditioned continuation.

Our component ablation (Table E.2\color{blue}{\text{Table E.2}}) isolates EKO while retaining maximin selection. Across 24 dataset–generator configurations, adding EKO improves SF and ASR in all settings, increasing average SF from 0.6679 to 0.7365 and ASR from 90.15% to 97.57%, while CR decreases from 87.65% to 84.71%. Nevertheless, harmonic utility improves from 0.7988 to 0.8411, with gains in 23/24 configurations.

These results establish a coverage–fidelity–acceptance trade-off, rather than uniform improvements. EKO enhances response fidelity and acceptance, while maximin selection increases coverage from 0.780 (Dual-Rand) to 0.847 (GE³). Thus, EKO and maximin selection serve distinct, complementary purposes: the former improves extraction fidelity and acceptance, while the latter promotes broader retrieval coverage within the dual-track framework. We have made this trade-off explicit in the revised Section 4.4\color{blue}{\text{Section 4.4}}.


Weakness 1(b), Questions 2(a) and 2(b). Metric Weighting and Sensitivity

We use equal-weight harmonic utility to summarize three complementary objectives—coverage, fidelity, and attack success—while penalizing poor performance in any one dimension.

To test whether equal weighting determines our conclusions, we exhaustively enumerate all nonnegative weight combinations for CR, SF, and ASR that sum to one, using increments of 0.05. As the CR weight increases from 0 to 1, the remaining two weights admit 21, 20, …, 1 allocations, yielding 231 combinations in total. Equal one-third weights are evaluated separately.

We report the results on a 0.05 grid under harmonic, arithmetic, and geometric aggregation (Table E.3\color{blue}{\text{Table E.3}}). GE³ strictly outperforms the other three variants under 191/231 (82.7%), 188/231 (81.4%), and 189/231 (81.8%) weight settings, respectively. It also ranks first under all three equal-weight formulations. Coverage-only weighting favors Ans-MMD, while SF-only and ASR-only weighting favor Dual-Rand.

These results demonstrate that GE³'s overall advantage extends across a broad range of evaluation priorities and does not depend solely on equal weighting, although specialized objectives may favor individual ablation variants.

AuthorsOpenReview ↗

Response to Reviewer x4YS [2/2]

Weakness 2, Question 2(c). Track Utilization and Complementarity

We measured the sources of executed queries across eight 200-request trajectories. Excluding seeds, 1,116/1,592 queries (70.1%) originate from EKO and 476/1,592 (29.9%) from answer-grounded exploration. Both tracks contribute in every trajectory, with EKO utilization ranging from 55.3% to 80.9%. The utilization results demonstrate that both tracks actively contribute to query selection.

The two tracks provide complementary sources of grounded candidates. Answer-grounded exploration generates verification and discovery questions from facts expressed in the primary answer, whereas EKO extracts follow-up cues directly from the victim-generated continuation span, potentially leveraging retrieved information absent from the primary answer. Rather than enforcing a predetermined balance between the two tracks, we merge their candidates into a shared pool and select queries based on their semantic novelty relative to recent query history. This allows both tracks to contribute according to the relevance and diversity of their candidates.


Question 1. Does the continuation suffix change retrieval?

Yes. Comparing the same 32 questions with and without the suffix, the average Top-10 overlap is 4.625 passages for BGE-M3 and 8.531 for Qwen3-Embedding-8B (Appendix B.2\color{blue}{\text{Appendix B.2}}). Thus, the suffix can alter retrieval, with the extent depending on the retriever.

Importantly, EKO operates on the actual retrieved context of the suffixed query and does not require retrieval invariance. To establish whether continuation cues depend on this context, we additionally conducted controlled passage-deletion and replacement experiments (Appendix B.2\color{blue}{\text{Appendix B.2}}). Under BGE-M3, source-specific five-word phrase matches across all four corpora decrease from 63/320 with complete context to 1/320 after source deletion, while replacement passages introduce 53/320 new-source matches. Crucially, holding the primary answer fixed preserves this pattern: original-source matches decrease from 55/320 to 5/320 after deletion, while replacement introduces 61/320 new-source matches.

These interventions provide strong evidence that continuation cues can depend on retrieved information absent from the primary answer, rather than merely reflecting changes in that answer. While retrieval shifts can also contribute to extraction, the controlled results directly support the context-dependent mechanism underlying EKO.


Question 3. Shadow RAG: Knowledge Beyond the Closed-Book LLM

We added a 100-question Healthcare evaluation to distinguish the contribution of extracted knowledge from the generator's pretrained knowledge (Section 4.3\color{blue}{\text{Section 4.3}}). All conditions use the same Llama-3.1-8B generator, and Shadow RAG relies exclusively on attacker-visible extracted answers collected from 1,000 target requests.

On 100 questions targeting facts contained in GE³'s extracted corpus and verified against original records, providing the extracted corpus increases necessary-fact completeness from 18/100 (closed-book) to 92/100 (Shadow RAG). Here, necessary facts are the key pieces of information that an answer must include, established from the source record before evaluation. Completeness requires all essential source-supported facts to be correctly conveyed. Since the generator remains identical, the improvement demonstrates the contribution of attacker-visible extracted knowledge beyond pretrained knowledge. We report the added evaluation in the revised Section 4.3\color{blue}{\text{Section 4.3}}.


Minor Comments. Terminology, ablation labels and grammar

We have corrected the 84.7% relative response rate terminology in the abstract, standardized ablation labels as Ans-Rand, Ans-MMD, Dual-Rand, and GE³, and corrected “Appendices A and H.”

We sincerely appreciate your time and effort in reviewing our paper. We have provided detailed responses to your comments, along with additional experiments and revisions to address your concerns.

As the discussion period is drawing to a close, we would be very grateful if you could take a moment to review our responses when convenient. If you have any remaining concerns or would like further clarification, we would be more than happy to address them.

Thank you again for your valuable feedback and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair, Senior Area Chairs, and Reviewers,

We sincerely thank all of you for the time and effort devoted to reviewing our paper and participating in the discussion. In our point-by-point rebuttals, we have addressed each question and concern raised by the reviewers, and we have revised the manuscript accordingly with additional experiments, analyses, and clarifications.

Regarding the methodological novelty of our work, we would like to clarify that GE³'s contribution lies not in applying standalone query-generation or diversity-selection heuristics, but in coupling dual-track knowledge-grounded query formation with history-relative exploration selection. Specifically, existing adaptive attacks generate follow-up queries solely from information already expressed in returned answers; in contrast, GE³ forms candidate queries grounded in two complementary layers of the target RAG system—deriving exploration queries from facts stated in the primary answer while exploiting the system's internal Echo Knowledge (Echo Knowledge Overflow, EKO), i.e., retrieved passages in the generation context that remain unexpressed in the primary answer and are surfaced through a benign continuation suffix. Because both tracks remain anchored in the underlying knowledge base, the candidate pool preserves high semantic fidelity and query acceptance while extending probing beyond answer-visible facts; building on this grounded pool, history-relative maximin selection then evaluates candidates against prior queries to identify the most promising direction for expanding corpus coverage without inducing out-of-domain semantic drift.

The reviewers recognized the practical relevance of the extraction threat (Reviewers x4YS\color{teal}{\text{x4YS}}, zo3e\color{olive}{\text{zo3e}}, and n4i5\color{purple}{\text{n4i5}}), GE³'s intuitive and coherent black-box framework (Reviewers x4YS\color{teal}{\text{x4YS}}, zo3e\color{olive}{\text{zo3e}}, and n4i5\color{purple}{\text{n4i5}}), and the solid experiments and ablations across four corpora and twelve LLMs (Reviewers x4YS\color{teal}{\text{x4YS}}, zo3e\color{olive}{\text{zo3e}}, and PDoU\color{brown}{\text{PDoU}}). In the follow-up discussion, Reviewer n4i5\color{purple}{\text{n4i5}} noted that EKO provides “an interesting source of retrieval-grounded information,” found the new mechanism and defense experiments convincing, and raised the overall Rating from 2 to 3; Reviewer zo3e\color{olive}{\text{zo3e}} stated that their “primary concerns have been well addressed” and raised Confidence (3 to 4) and Informativeness (2 to 3, retaining Rating 3); and Reviewers x4YS\color{teal}{\text{x4YS}} and PDoU\color{brown}{\text{PDoU}} have not yet posted follow-up responses.

Thank you once again for your time, feedback, and consideration.

Best regards,

Authors of Submission 27530

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.