acceptodds
Under review as a conference paper at ICLR 2027

TopoPoison: Topological Poisoning Attacks on Graph Retrieval-Augmented Generation

Abstract

Graph Retrieval-Augmented Generation (GraphRAG) enhances the ability of large language models to integrate knowledge across documents and answer multi-hop questions by organizing unstructured corpora into entities, relations and higher-level graph structures. However, its continued reliance on external knowledge sources also introduces new data poisoning risks. Existing poisoning methods developed for conventional RAG do not readily transfer to GraphRAG, as its structured processing transforms injected text into graph representations and changes how poisoned knowledge is retrieved. In GraphRAG, successful targeted poisoning involves three successive challenges: preserving the poisoned knowledge during structured indexing, ensuring its retrieval for the target query, and inducing the attacker-specified answer despite competing evidence during answer generation. Black-box settings further compound these challenges by limiting the attacker's visibility into the internal knowledge graph and retrieval structures, making it more difficult to identify and manipulate query-relevant reasoning structures. To address these challenges, we propose TopoPoison, a black-box, append-only targeted data poisoning framework for GraphRAG. TopoPoison follows a three-stage pipeline: it first recovers an approximate query-relevant graph from victim feedback, organizes recovered reasoning paths into candidate-specific subgraphs, and estimates the structural support of each candidate; it then uses these structures to reinforce the attacker-specified answer and redirect selected competitors based on the estimated structural support; finally, it realizes the planned structural interventions as coherent poisoning documents with high indexing fidelity and retrieval reachability. Experimental results show that TopoPoison consistently outperforms representative GraphRAG poisoning baselines across different datasets and victim systems, demonstrating the feasibility of black-box targeted data poisoning using only query feedback and revealing potential poisoning risks in the external knowledge ingestion process of GraphRAG.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Security, Retrieval-Augmented Generation (RAG), Graph Retrieval-Augmented Generation (GraphRAG), Data Poisoning, Adversarial Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EsUzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies targeted data poisoning attacks on Graph Retrieval-Augmented Generation (GraphRAG) systems and proposes TopoPoison, a black-box attack framework. The core idea is to treat the reasoning subgraph for each candidate answer as the basic unit for poisoning intervention. The full pipeline runs in three stages: query structure recovery, candidate support manipulation, and poisoning text realization, and it works purely using response feedback from the victim, with no access to internal data. Experiments across 3 multi-hop QA benchmarks and 2 mainstream GraphRAG systems show the method clearly outperforms existing baselines like GragPoison and KEPo on targeted attack success rate.

Critical strengths
  1. The reasoning subgraph corresponding to the candidate answer is used as the basic unit of poisoning intervention, and the attack strategy is designed from the perspective of topological structure support, which has positive reference significance for understanding the structural vulnerability of GraphRAG.
  2. The experimental setup is thorough. It covers multiple datasets, and includes component ablations, budget sensitivity analysis, cross-model transferability tests, and impact measurements on non-target queries.
Critical weaknesses
  1. All experiments run on just two systems: MS-GraphRAG and LightRAG. Those are both fairly classic implementations: one uses hierarchical community summaries, the other mixes graph and vector retrieval. But the field has moved on: there are newer state-of-the-art methods like GFM-RAG, HippoRAG 2, and G-Reasoner that use graph foundation models, personalized PageRank for associative retrieval, structured multi-hop reasoning chains, and so on. These have very different graph construction and retrieval mechanics, and very different topological properties. The paper never tests on any of these more advanced architectures, or analyzes how different graph-building strategies affect poisoning robustness.
  2. This paper sets the attack as a purely reactive black box attack, but the query structure recovery stage requires an average of nearly 20 relevant queries to the victim, which is easily recognized by access control and anomaly detection mechanisms in real deployment scenarios, and the concealment and practicability of the attack are in doubt.
  3. Although the paper is named "topology poisoning", the core intervention methods are still traditional association enhancement, fact negation and answer replacement. The utilization of graph topology only stops at the level of subgraph scoring and path ranking, and does not explore the unique attack points of graph structure (such as bridge nodes, community hubs, reasoning bottlenecks, etc.). Compared with the existing GraphRAG poisoning methods, the core differences are mainly in the granularity of attack organization, the underlying poisoning logic has no essential innovation, and the technical increment is limited.
Questions for authors
  1. In the real scenario, high-frequency queries on the same topic are easy to trigger anomaly detection. How can the author evaluate the concealment and feasibility of this attack in the actual deployment? Are attack performance results under low query budget provided?
  2. The candidate support score is a heuristic on the attacker side, not aligned with the victim’s internal retrieval ranking logic. How do you know the budget allocation strategy based on this score is actually optimal? Could scoring bias lead to misallocation of poisoning resources?
  3. Experiments only cover two classic GraphRAG architectures. Why did you not include more recent state-of-the-art systems like GFM-RAG or HippoRAG 2 in your evaluation? Do you expect this attack to generalize equally well across graph architectures with very different topological properties?
8 comments
AuthorsOpenReview ↗

W1 & Q3 — Evaluation on Additional GraphRAG Architectures.

We thank the reviewer for suggesting a broader evaluation.

We extended our experiments to HippoRAG 2 and GFM-RAG, using the G-reasoner-34M configuration for the latter. These systems extend our evaluation beyond community-based and hybrid graph–vector retrieval: HippoRAG 2 uses Personalized PageRank for associative retrieval, whereas the evaluated GFM-RAG configuration uses a pretrained graph foundation model for graph reasoning and retrieval. G-reasoner-34M serves as the graph model, with a separate LLM responsible for answer generation. On MuSiQue, we compared TopoPoison against GragPoison, KEPo-S, and KEPo-M. The results are reported below.

Table 1: Effectiveness of GraphRAG poisoning attacks against HippoRAG 2 on MuSiQue.

MethodTASRC-TASRUASRC-UASR
GragPoison46.0%43.3%74.0%61.7%
KEPo-S65.0%65.0%79.0%68.3%
KEPo-M66.0%65.0%82.0%71.7%
TopoPoison84.0%88.3%92.0%91.7%

Table 2: Effectiveness of GraphRAG poisoning attacks against GFM-RAG (with G-reasoner-34M) on MuSiQue.

MethodTASRC-TASRUASRC-UASR
GragPoison22.0%18.6%57.0%37.3%
KEPo-S52.0%49.2%74.0%57.6%
KEPo-M62.0%59.3%75.0%62.7%
TopoPoison78.0%71.2%89.0%81.4%

Cross-system interpretation. TopoPoison outperformed all evaluated baselines on all four metrics on both systems. Its TASR reached 84.0% on HippoRAG2 and 78.0% on GFM-RAG exceeding the best baseline on each system by 18 and 16 percentage points, respectively.

These results show that TopoPoison retains its targeted attack advantage on two architectures with different retrieval mechanisms. We have added the new results, their experimental configurations and this interpretation to Section 5 and Appendix E.

AuthorsOpenReview ↗

W2 & Q1 — Query Budget and Deployment Feasibility

We thank the reviewer for highlighting the practical implications of repeated victim queries. Although Table 17 in the revised manuscript reports averages of 18.5 and 19.4 queries per target question on MS-GraphRAG and LightRAG, respectively, these interaction costs are not prerequisites for effective poisoning under the evaluated setting. Our additional experiments show that a single-query variant achieves TASRs of 61.0% and 85.0%, compared with 63.0% and 85.0% under the full query budget. We detail the experimental settings and discuss the implications for deployment feasibility below.

(1) Attack performance under limited query budgets.

We evaluated TopoPoison on MuSiQue with different limits on the number of victim queries per target question, denoted by BqB_{q}. This budget includes the initial interaction and subsequent follow-up queries used for structure recovery. Whereas the default recovery procedure uses separate queries to obtain entity and relation information, the single-query variant Bq=1B_{q}=1 obtains both through one victim interaction. All remaining settings were kept unchanged. The following tables report the query limit, the average number of queries actually issued, and the resulting attack performance.

Table 3: Effect of the victim query budget against MS-GraphRAG on MuSiQue.

BqB_qAvg. QueriesTASRC-TASRUASRC-UASR
11.0061.0%61.7%85.0%74.5%
22.0062.0%59.6%82.0%66.0%
65.7661.0%63.8%87.0%74.5%
109.3261.0%55.3%81.0%59.6%
1412.0063.0%55.3%84.0%68.1%
1813.9864.0%61.7%85.0%70.2%
2215.3961.0%63.8%86.0%72.3%
2616.3965.0%57.4%83.0%63.8%
3017.2765.0%61.7%83.0%66.0%
3417.9563.0%57.4%81.0%61.7%
3818.4761.0%53.2%80.0%59.6%
4218.5063.0%61.7%82.0%63.8%

Table 4: Effect of the victim query budget against LightRAG on MuSiQue.

BqB_qAvg. QueriesTASRC-TASRUASRC-UASR
11.0085.0%82.5%92.0%86.0%
22.0087.0%84.2%93.0%89.5%
65.4985.0%82.5%94.0%89.5%
108.3687.0%86.0%93.0%89.5%
1410.7687.0%86.0%95.0%91.2%
1812.6386.0%89.5%93.0%91.2%
2214.2586.0%84.2%92.0%87.7%
2616.1784.0%82.5%94.0%89.5%
3017.3085.0%86.0%93.0%87.7%
3418.2388.0%86.0%94.0%89.5%
3819.0386.0%87.7%94.0%91.2%
4219.4085.0%86.0%91.0%91.2%

TASR ranges from 61.0% to 65.0% on MS-GraphRAG and from 84.0% to 88.0% on LightRAG across the tested budgets, with no consistent upward trend. Reducing BqB_q from 42 to 1 changes TASR by −2.0 and 0.0 percentage points, respectively, while reducing the average interactions from 18.5 and 19.4 to one.

The reason is the limited scope of the intervention. Under the approximately 300-word poisoning budget, TopoPoison reinforces the target answer and redirects at most one competing-answer subgraph K=1K=1. In this setting, a small number of queries can recover enough local structure to plan the intervention. A larger query budget permits more extensive exploration, but the present results do not establish that such exploration is necessary under this constrained intervention budget.

AuthorsOpenReview ↗

W2 & Q1 — Query Budget and Deployment Feasibility (continued)

(2) Implications for deployment feasibility and detectability.

We agree that repeated, closely related queries could trigger rate limits or anomaly detection in deployed systems. TopoPoison’s follow-up queries seek clarification about entities, relations, or intermediate facts identified in earlier responses, and can therefore resemble ordinary clarification requests.

Prior studies have shown that user-visible GraphRAG responses can expose entity and relation information [1], and that adaptive follow-up queries can support incremental graph recovery [2]. These findings motivate feedback-based recovery, while our objective is narrower: recovering local information relevant to a target question for subsequent poisoning. The single-query results provide direct evidence that, for the evaluated setting, this objective can be achieved with substantially fewer interactions than those reported for our default recovery procedure. They address the concern that the attack necessarily depends on repeated querying, while leaving its detectability under specific deployment controls as an open empirical question.

The results address the interaction-cost concern in the evaluated setting; concealment remains an open empirical question. We have added the budget sweep to Appendix E and stated the deployment scope in the Limitations.

References

[1] Jiale Liu, Jiahao Zhang, and Suhang Wang. 2026. Exposing Privacy Risks in Graph Retrieval-Augmented Generation. In Findings of the Association for Computational Linguistics: ACL 2026, pages 18073–18093, San Diego, California, United States. Association for Computational Linguistics.

[2] Shuhua Yang, Jiahao Zhang, Yilong Wang, Dongwon Lee, and Suhang Wang. 2026. Query-Efficient Agentic Graph Extraction Attacks on GraphRAG Systems. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16017–16041, San Diego, California, United States. Association for Computational Linguistics.

Reviewer_u2mnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopoPoison studies append-only targeted poisoning against GraphRAG under black-box access. It recovers query-relevant entities and relations from victim feedback, groups paths by candidate answer, and generates documents that reinforce an attacker-specified answer while redirecting competing support. Across three multi-hop QA datasets and two GraphRAG implementations, the reported TASR gains over the strongest baseline range from 13.0 to 38.0 percentage points. The problem is relevant, and the end-to-end improvements are substantial. However, unequal query budgets, the large effect of file organization, and limited validation of the recovered support structures make it difficult to attribute these gains specifically to the proposed topology-based planning.

Critical strengths
  1. The threat model clearly distinguishes response-only access from evidence-visible access and restricts the attacker to adding documents. Separating indexing, retrieval, and answer generation also provides a useful framework for studying where poisoning succeeds or fails.

  2. The evaluation covers three datasets and two GraphRAG systems, and the main improvements are consistent across all six settings. Reporting C-TASR is particularly useful because it separates targeted corruption of initially correct answers from attacks on queries the victim already answers incorrectly.

  3. The appendices provide substantial supporting analysis, including recovery quality, candidate selection, text organization, attack cost, collateral effects, and index-preserving chat-model changes. The paper generally qualifies these findings carefully, including acknowledging that target presence does not establish relation preservation or a causal mechanism.

Critical weaknesses
  1. The main comparison does not control for the information acquired through victim interaction. Table 11 reports 18.5–19.4 victim queries per target for TopoPoison, versus 0–1 for the baselines. Matching injected word counts therefore does not establish an advantage under comparable attack resources. This matters because additional feedback can improve both candidate discovery and the specificity of the generated misinformation. File organization is another potentially substantial confound: Table 10 shows that storing identical texts separately improves TASR by 31 points on MS-GraphRAG and 12 points on LightRAG relative to ordered concatenation. The paper should establish whether every baseline receives comparable file organization and evaluate performance under matched query budgets.

  2. The experiments do not adequately isolate the contribution of candidate-specific topology from the surrounding text-generation pipeline. In Table 2, the topology-free variant already achieves 56% and 79% TASR, compared with 63% and 85% for the full method. Removing redirection alone changes LightRAG TASR by only one point. Appendix D.3 implements ablations by removing prompt requirements and expanding or polishing the remaining text, which also changes wording, repetition, and potentially retrieval relevance. These results support the combined pipeline, but provide weaker evidence that subgraph construction and support ranking account for its large advantage over prior attacks.

  3. Some retained “supporting paths” do not establish support for the answer relation required by the query. In Appendix H, four-hop paths connect different markers through Kidney Dysfunction. The appendix explicitly acknowledges that this connectivity does not establish that two medications share the same marker, yet these paths contribute to the candidate-support score in Table 18. This creates a mismatch with Section 4.2's requirement to retain only coherent supporting paths. Moreover, Equation 5 sums overlapping paths, allowing several routes through shared evidence to increase a candidate's score without adding independent support. Although the score is explicitly heuristic, its interpretation and semantic acceptance criteria need stronger justification.

  4. The proposed indexing-fidelity mechanism is not directly validated against the victim's extracted relations. Equation 12 checks coverage using a local extractor, whereas the stage-wise experiments count normalized occurrences of the target name in index files and retrieved context. These measurements cannot determine whether the intended predicates, negations, temporal qualifications, or attachment points survive victim indexing. Table 8 also reports relatively low relation F1, between 0.203 and 0.262. The paper acknowledges these limitations, but the evidence still leaves open whether structural intervention, repeated target exposure, or persuasive knowledge-update wording drives the observed success.

  5. Uncertainty and evaluator reliability are insufficiently characterized. Each main setting contains 100 targets, and the clean-correct subsets contain only 46–88 queries. Several component and candidate-selection differences therefore correspond to very few outcomes. The paper does not report uncertainty intervals or repeated end-to-end runs; the three random candidate selections address only one source of variation. Attack success and clean correctness are judged by deepseek-v3.2 without a reported human agreement study. This does not invalidate the large main gains, but it limits confidence in the smaller design conclusions and in how ambiguous, multi-answer, or internally contradictory responses are classified.

Questions for authors
  1. Report TASR against victim-query budget, including one-query and intermediate-budget TopoPoison variants. Allow baselines to use the same feedback or provide a control that generates poisoning text directly from the same recovered transcript without constructing candidate subgraphs.

  2. Apply separate-file, ordered-concatenation, and shuffled-concatenation ingestion consistently across methods. State the file and chunk counts for each baseline so that the large organization effect can be separated from the attack-planning effect.

8 comments
AuthorsOpenReview ↗

W1 & Q2 — Query Budget and File Organization

We thank the reviewer for identifying these potential confounding factors.

We conducted additional experiments to examine both victim-query budgets and file organization. With TopoPoison restricted to one victim query per target question and all methods evaluated under the same file-organization rules, TopoPoison retained the highest TASR on both systems. Under separate-file ingestion, it exceeded the strongest baseline by 10 and 11 percentage points on MS-GraphRAG and LightRAG, respectively. File-level isolation is a design component of TopoPoison, and this comparison tests all methods with access to it. We describe the experimental controls and results below.

(1) Query budget.

We evaluated TopoPoison on MuSiQue with different limits on the number of victim queries per target question, denoted by BqB_{q}. This budget includes the initial interaction and subsequent follow-up queries used for structure recovery. Whereas the default recovery procedure uses separate queries to obtain entity and relation information, the single-query variant Bq=1B_{q}=1 obtains both through one victim interaction. All remaining settings were kept unchanged. The following tables report the query limit, the average number of queries actually issued, and the resulting attack performance.

Table 3: Effect of the victim query budget against MS-GraphRAG on MuSiQue.

BqB_qAvg. QueriesTASRC-TASRUASRC-UASR
11.0061.0%61.7%85.0%74.5%
22.0062.0%59.6%82.0%66.0%
65.7661.0%63.8%87.0%74.5%
109.3261.0%55.3%81.0%59.6%
1412.0063.0%55.3%84.0%68.1%
1813.9864.0%61.7%85.0%70.2%
2215.3961.0%63.8%86.0%72.3%
2616.3965.0%57.4%83.0%63.8%
3017.2765.0%61.7%83.0%66.0%
3417.9563.0%57.4%81.0%61.7%
3818.4761.0%53.2%80.0%59.6%
4218.5063.0%61.7%82.0%63.8%

Table 4: Effect of the victim query budget against LightRAG on MuSiQue.

BqB_qAvg. QueriesTASRC-TASRUASRC-UASR
11.0085.0%82.5%92.0%86.0%
22.0087.0%84.2%93.0%89.5%
65.4985.0%82.5%94.0%89.5%
108.3687.0%86.0%93.0%89.5%
1410.7687.0%86.0%95.0%91.2%
1812.6386.0%89.5%93.0%91.2%
2214.2586.0%84.2%92.0%87.7%
2616.1784.0%82.5%94.0%89.5%
3017.3085.0%86.0%93.0%87.7%
3418.2388.0%86.0%94.0%89.5%
3819.0386.0%87.7%94.0%91.2%
4219.4085.0%86.0%91.0%91.2%

TASR ranges from 61.0% to 65.0% on MS-GraphRAG and from 84.0% to 88.0% on LightRAG across the tested budgets, with no consistent upward trend. Reducing BqB_q from 42 to 1 changes TASR by −2.0 and 0.0 percentage points, respectively, while reducing the average interactions from 18.5 and 19.4 to one.

The reason is the limited scope of the intervention. Under the approximately 300-word poisoning budget, TopoPoison reinforces the target answer and redirects at most one competing-answer subgraph K=1K=1. In this setting, a small number of queries can recover enough local structure to plan the intervention. A larger query budget permits more extensive exploration, but the present results do not establish that such exploration is necessary under this constrained intervention budget.

(2) File organization.

The original comparison used method-specific ingestion: GragPoison followed the shuffling option of its public implementation, our KEPo reproduction used ordered concatenation because the paper did not specify file organization and no public implementation was available, and TopoPoison stored each text separately. File-level isolation is part of TopoPoison's Stage 3 design (Figure 2D). This design stems from our observation that existing GraphRAG systems typically use documents as boundaries when chunking. We therefore organize each planned intervention as a separate poisoning document, aiming to reduce the mixing of different intervention contexts during chunking.

AuthorsOpenReview ↗

W1 & Q2 — Query Budget and File Organization (continued)

(2) File organization. (continued)

We evaluated all four methods on the same 100 MuSiQue targets under Ordered, Shuffled and Separate ingestion. All methods used at most one victim query per target question: GragPoison used no victim queries, KEPo-S and KEPo-M each used one, and TopoPoison used its single-query variant. Both systems used a chunk size of 1,200 tokens and an overlap of 100 tokens, with about 300 injected words per target and at most one victim query per target. To make injection volumes and file counts comparable, each baseline text was duplicated once with light edits; each method's resulting text set was then held fixed across organizations. Under Separate, TopoPoison, KEPo-S and KEPo-M each used 200 files and 200 chunks; GragPoison used 194 files and 194 chunks, corresponding to 97 selected relations shared across targets. Under Ordered and Shuffled, each method used one file, producing 41–42 chunks on MS-GraphRAG and 40–43 on LightRAG.

Table 1: Attack effectiveness under consistent file organization against MS-GraphRAG on MuSiQue.

OrganizationMethodTASRC-TASRUASRC-UASR
OrderedGragPoison17.0%12.8%64.0%31.9%
OrderedKEPo-S29.0%23.4%64.0%27.7%
OrderedKEPo-M28.0%25.5%68.0%38.3%
OrderedTopoPoison32.0%25.5%69.0%38.3%
ShuffledGragPoison20.0%10.6%60.0%27.7%
ShuffledKEPo-S29.0%25.5%66.0%31.9%
ShuffledKEPo-M23.0%25.5%65.0%31.9%
ShuffledTopoPoison34.0%31.9%70.0%40.4%
SeparateGragPoison39.0%38.3%81.0%61.7%
SeparateKEPo-S49.0%51.1%78.0%57.4%
SeparateKEPo-M51.0%46.8%75.0%51.1%
SeparateTopoPoison61.0%61.7%85.0%74.5%

Table 2: Attack effectiveness under consistent file organization against LightRAG on MuSiQue.

OrganizationMethodTASRC-TASRUASRC-UASR
OrderedGragPoison56.0%49.1%73.0%57.9%
OrderedKEPo-S53.0%47.4%67.0%49.1%
OrderedKEPo-M46.0%35.1%59.0%40.4%
OrderedTopoPoison75.0%75.4%86.0%77.2%
ShuffledGragPoison47.0%38.6%65.0%50.9%
ShuffledKEPo-S53.0%50.9%69.0%54.4%
ShuffledKEPo-M54.0%43.9%67.0%47.4%
ShuffledTopoPoison69.0%70.2%86.0%78.9%
SeparateGragPoison61.0%54.4%82.0%75.4%
SeparateKEPo-S74.0%68.4%84.0%73.7%
SeparateKEPo-M66.0%59.6%78.0%63.2%
SeparateTopoPoison85.0%82.5%92.0%86.0%

Organization changes every method's results substantially; for example, KEPo-S on MS-GraphRAG rises from 29.0% under ordered concatenation to 49.0% under separate files. TopoPoison has the highest TASR under all three schemes in both systems. Under Separate it exceeds the strongest baseline by 10 and 11 percentage points (61.0% vs 51.0% on MS-GraphRAG; 85.0% vs 74.0% on LightRAG). Under ordered concatenation on MS-GraphRAG the margin is 3.0 percentage points, and under shuffled concatenation 5.0 percentage points.

These findings suggest a potential trade-off in GraphRAG document organization. Keeping documents separate may help preserve coherent poisoning content during indexing, while mixing unrelated topics may introduce ambiguity into entity and relation extraction and affect clean-task performance. GraphRAG designers should therefore evaluate document organization jointly in terms of poisoning robustness, indexing accuracy, and answer quality.

The controlled results support an advantage for the combined TopoPoison pipeline under comparable query limits and ingestion rules. File-level isolation remains an explicit component of the method, and the results show that it also benefits the baselines. We have incorporated these results and the corresponding experimental details into the revised manuscript.

AuthorsOpenReview ↗

W2 & W5 — Isolating topology; uncertainty and evaluator reliability

We thank the reviewer for raising these concerns. We strengthened the evaluation in two ways: text-level ablations on an expanded set of 300 target questions, and human validation of the model-based judgments.

(1) Text-level ablations and expanded evaluation.

To address potential confounds introduced by regenerating poisoning texts, we conducted additional ablations by directly editing the texts produced by the full method. For each ablation, we deleted the passages corresponding to the removed component and left the remaining content unchanged. This avoids rewriting or polishing the retained text, thereby controlling changes in its wording.

To alleviate uncertainty, We also expanded the evaluation to 300 target questions, randomly selected from the 500-question MuSiQue pool used to construct the clean corpus. We followed the same proportional stratified sampling procedure by reasoning hop count as in the main experiments. We evaluated the full method and its component ablations on MS-GraphRAG and LightRAG, and compared the four candidate-selection strategies on the same questions.

Component ablations. The following tables summarize the results. w/o Redir. removes competitor redirection, w/o Rein. removes target reinforcement, w/o Topology removes both, and w/o Q–E Conn. removes the explicit connections between query entities and the target answer.

Table 3: Effect of component ablation against MS-GraphRAG on the MuSiQue.

SettingTASRC-TASRUASRC-UASR
Full58.7%60.1%84.7%68.2%
w/o Redir.52.0%58.2%81.3%62.5%
w/o Rein.49.3%52.7%82.0%61.6%
w/o Topology53.3%54.9%83.0%65.4%
w/o Q–E Conn.51.7%51.9%80.3%59.2%

Table 4: Effect of component ablation against LightRAG on the MuSiQue.

SettingTASRC-TASRUASRC-UASR
Full86.0%88.8%94.0%91.0%
w/o Redir.75.7%77.2%88.7%83.0%
w/o Rein.82.0%83.7%91.0%87.7%
w/o Topology81.3%81.9%90.3%85.3%
w/o Q–E Conn.84.3%85.2%93.7%88.6%

The full configuration achieved the highest mean TASR on both systems. Removing both target reinforcement and competitor redirection reduced mean TASR from 58.7% to 53.3% on MS-GraphRAG and from 86.0% to 81.3% on LightRAG. Removing query–entity connections also reduced TASR on both systems. These comparisons provide evidence that the intervention content generated by structural planning and the explicit query–entity connections contribute to targeted attack success.

The effects varied across systems and ablations. For example, removing competitor redirection reduced mean TASR from 86.0% to 75.7% on LightRAG. We therefore interpret the results as evidence for the contributions of these components within the combined poisoning pipeline, rather than assuming that their effects are uniform or additive.

Candidate selection. We additionally compared selecting the highest-scoring candidate (Best), the victim’s original answer (Default), a random candidate (Random), and the lowest-scoring candidate (Lowest).

Table 5: Effect of candidate selection against MS-GraphRAG on MuSiQue.

SettingTASRC-TASRUASRC-UASR
Best58.7%60.1%84.7%68.2%
Default57.3%62.2%86.0%71.9%
Random57.0%57.7%84.0%66.7%
Lowest55.3%60.0%84.7%69.0%

Table 6: Effect of candidate selection against LightRAG on MuSiQue.

SettingTASRC-TASRUASRC-UASR
Best86.0%88.8%94.0%91.0%
Default84.7%87.2%94.3%90.6%
Random83.7%85.2%93.0%89.4%
Lowest82.0%83.1%92.3%87.6%

Highest-score selection achieved the highest mean TASR on both systems: 58.7% and 86.0%, compared with 57.3% and 84.7% for selecting the victim’s original answer. The improvement is modest, indicating that the original answer is itself a competitive intervention target. These results provide preliminary empirical support for support-based candidate prioritization. TASR is the primary criterion for this comparison because the objective is to induce the attacker-specified answer, rather than an arbitrary incorrect response.

Reviewer_DtspOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopoPoison studies targeted, append-only poisoning of GraphRAG systems under response-only access. It queries the victim to construct an approximate graph of candidate answers and supporting relations, ranks candidate-specific subgraphs, and generates documents that reinforce the chosen false answer and redirect support away from competing answers. Experiments on three question-answering datasets and two GraphRAG systems report large gains over GragPoison and two KEPo variants. The paper also studies intermediate target presence, component ablations, document organization, poisoning volume, and limited defenses.

Critical strengths
  1. The threat and success criteria are very useful. This helps separate induced errors from questions the victim already answered incorrectly

  2. The paper provide thorough details in appendix, which contains full prompt catalog, exact victim configurations and versions, dataset construction with seed.

Critical weaknesses
  1. Table 11 reports 18.5 and 19.4 victim queries per target for TopoPoison, versus 0–1 for the baselines. Matching injected word counts therefore does not isolate the contribution of candidate-level topology from the addtional victim feedback.

  2. The baseline description does not clearly establish a comparison in which all methods use the same file boundaries and chunking conditions.

Questions for authors
  1. Were baseline poisoning texts ingested with the same separate-file organization as TopoPoison?

  2. For each dataset–victim setting, how many of the 100 target queries produced no valid supporting path and therefore no manipulation plan?

  3. How many files, and of what size distribution, constitute each clean corpus?

8 comments
AuthorsOpenReview ↗

W1 - Victim feedback and the contribution of candidate-level topology

We thank the reviewer for raising this concern. TopoPoison retains similar TASR under a single-query limit. This addresses the extra-interaction confound; the contribution of candidate-level topology requires the component comparisons as a separate test.

(1) Sensitivity to the victim-query budget.

We evaluated TopoPoison on MuSiQue with different limits on the number of victim queries per target question, denoted by BqB_{q}. This budget includes the initial interaction and subsequent follow-up queries used for structure recovery. Whereas the default recovery procedure uses separate queries to obtain entity and relation information, the single-query variant Bq=1B_{q}=1 obtains both through one victim interaction. All remaining settings were kept unchanged. The following tables report the query limit, the average number of queries actually issued, and the resulting attack performance.

Table 1: Effect of the victim query budget against MS-GraphRAG on MuSiQue.

BqB_qAvg. QueriesTASRC-TASRUASRC-UASR
11.0061.0%61.7%85.0%74.5%
22.0062.0%59.6%82.0%66.0%
65.7661.0%63.8%87.0%74.5%
109.3261.0%55.3%81.0%59.6%
1412.0063.0%55.3%84.0%68.1%
1813.9864.0%61.7%85.0%70.2%
2215.3961.0%63.8%86.0%72.3%
2616.3965.0%57.4%83.0%63.8%
3017.2765.0%61.7%83.0%66.0%
3417.9563.0%57.4%81.0%61.7%
3818.4761.0%53.2%80.0%59.6%
4218.5063.0%61.7%82.0%63.8%

Table 2: Effect of the victim query budget against LightRAG on MuSiQue.

BqB_qAvg. QueriesTASRC-TASRUASRC-UASR
11.0085.0%82.5%92.0%86.0%
22.0087.0%84.2%93.0%89.5%
65.4985.0%82.5%94.0%89.5%
108.3687.0%86.0%93.0%89.5%
1410.7687.0%86.0%95.0%91.2%
1812.6386.0%89.5%93.0%91.2%
2214.2586.0%84.2%92.0%87.7%
2616.1784.0%82.5%94.0%89.5%
3017.3085.0%86.0%93.0%87.7%
3418.2388.0%86.0%94.0%89.5%
3819.0386.0%87.7%94.0%91.2%
4219.4085.0%86.0%91.0%91.2%

TASR ranges from 61.0% to 65.0% on MS-GraphRAG and from 84.0% to 88.0% on LightRAG across the tested budgets, with no consistent upward trend. Reducing BqB_q from 42 to 1 changes TASR by −2.0 and 0.0 percentage points, respectively, while reducing the average interactions from 18.5 and 19.4 to one.

The reason is the limited scope of the intervention. Under the approximately 300-word poisoning budget, TopoPoison reinforces the target answer and redirects at most one competing-answer subgraph K=1K=1. In this setting, a small number of queries can recover enough local structure to plan the intervention. A larger query budget permits more extensive exploration, but the present results do not establish that such exploration is necessary under this constrained intervention budget.

AuthorsOpenReview ↗

W1 - Victim feedback and the contribution of candidate-level topology (continued)

(2) Structural contribution.

We examine whether TopoPoison’s advantage persists when baselines receive the same recovered information and use component ablations to assess the contribution of target reinforcement and competitor redirection to targeted attack success.

Shared feedback. For each target question, we provided the baselines with the same entities and relations recovered from victim-query feedback for TopoPoison. GragPoison used this information to identify query-relevant entities and relations during relation selection, while KEPo-S and KEPo-M used it to identify anchor facts. We retained the remaining stages of each baseline’s pipeline.

In the tables below, Original denotes each baseline’s original setting, and Shared feedback denotes the setting augmented with the recovered entities and relations. The shared input is the extracted entity–relation information, rather than TopoPoison’s subsequent candidate-selection or intervention plans.

Table 3: Attack effectiveness with shared query feedback against MS-GraphRAG on MuSiQue.

MethodTASRC-TASRUASRC-UASR
GragPoison (Original)20.0%10.6%60.0%27.7%
GragPoison (Shared feedback)19.0%14.9%66.0%36.2%
KEPo-S (Original)25.0%19.1%62.0%23.4%
KEPo-S (Shared feedback)24.0%21.3%62.0%29.8%
KEPo-M (Original)20.0%17.0%62.0%25.5%
KEPo-M (Shared feedback)24.0%17.0%64.0%31.9%
TopoPoison63.0%61.7%82.0%63.8%

Table 4: Attack effectiveness with shared query feedback against LightRAG on MuSiQue.

MethodTASRC-TASRUASRC-UASR
GragPoison (Original)47.0%38.6%65.0%50.9%
GragPoison (Shared feedback)55.0%56.1%80.0%71.9%
KEPo-S (Original)54.0%49.1%72.0%59.6%
KEPo-S (Shared feedback)54.0%47.4%69.0%52.6%
KEPo-M (Original)51.0%43.9%67.0%52.6%
KEPo-M (Shared feedback)58.0%50.9%70.0%52.6%
TopoPoison85.0%86.0%91.0%91.2%

TopoPoison’s TASR exceeded that of the strongest baseline using shared feedback by 39 percentage points on MS-GraphRAG and 27 percentage points on LightRAG. Comparisons between each baseline’s original and shared-feedback settings measure the effect of adding feedback under its existing file organization. The separate experiment in W2&Q1 controls organization across methods.

Component ablation. In Table 12 of the revised manuscript, removing both target reinforcement and competitor redirection (w/o Topology) reduced TASR from 63.0% to 56.0% on MS-GraphRAG and from 85.0% to 79.0% on LightRAG. We also conducted a text-level ablation on 300 MuSiQue target questions, directly deleting the corresponding passages from the full method’s poisoning documents while preserving the remaining text (Reviewer u2mn, W2 & W5). TASR fell from 58.7% to 53.3% on MS-GraphRAG and from 86.0% to 81.3% on LightRAG. The consistent decreases indicate that the intervention content produced by candidate-level topology planning contributes to targeted attack success within the full pipeline.

These results show that TopoPoison remains effective with a single victim query, retains its advantage under shared feedback, and benefits from topology-guided interventions. We have added the query-budget and shared-feedback results and the corresponding experimental settings in the revised manuscript.

AuthorsOpenReview ↗

W2 & Q1 — Same file boundaries and chunking for all methods

We thank the reviewer for raising these concerns. In the original comparison, we used method-specific settings: we followed GragPoison's text shuffling strategy, used ordered concatenation in our KEPo reproduction because its paper did not specify a file organization scheme and a public implementation was unavailable; and stored TopoPoison's texts in separate files as part of its design (Figure 2D). In the new comparison, we apply the same three organization rules (Ordered, Shuffled, and Separate) and the same chunking settings to every method. TopoPoison achieves the highest TASR under each organization rule on both systems.

(1) Comparison under consistent file organization and query limits.

The original comparison used method-specific ingestion: GragPoison followed the shuffling option of its public implementation, our KEPo reproduction used ordered concatenation because the paper did not specify file organization and no public implementation was available, and TopoPoison stored each text separately. File-level isolation is part of TopoPoison's Stage 3 design (Figure 2D). This design stems from our observation that existing GraphRAG systems typically use documents as boundaries when chunking. We therefore organize each planned intervention as a separate poisoning document, aiming to reduce the mixing of different intervention contexts during chunking.

We evaluated all four methods on the same 100 MuSiQue targets under Ordered, Shuffled and Separate ingestion. Both systems used a chunk size of 1,200 tokens and an overlap of 100 tokens, with about 300 injected words per target and at most one victim query per target. To make injection volumes and file counts comparable, each baseline text was duplicated once with light edits; each method's resulting text set was then held fixed across organizations. Under Separate, TopoPoison, KEPo-S and KEPo-M each used 200 files and 200 chunks; GragPoison used 194 files and 194 chunks, corresponding to 97 selected relations shared across targets. Under Ordered and Shuffled, each method used one file, producing 41–42 chunks on MS-GraphRAG and 40–43 on LightRAG.

Table 1: Attack effectiveness under consistent file organization against MS-GraphRAG on MuSiQue.

OrganizationMethodTASRC-TASRUASRC-UASR
OrderedGragPoison17.0%12.8%64.0%31.9%
OrderedKEPo-S29.0%23.4%64.0%27.7%
OrderedKEPo-M28.0%25.5%68.0%38.3%
OrderedTopoPoison32.0%25.5%69.0%38.3%
ShuffledGragPoison20.0%10.6%60.0%27.7%
ShuffledKEPo-S29.0%25.5%66.0%31.9%
ShuffledKEPo-M23.0%25.5%65.0%31.9%
ShuffledTopoPoison34.0%31.9%70.0%40.4%
SeparateGragPoison39.0%38.3%81.0%61.7%
SeparateKEPo-S49.0%51.1%78.0%57.4%
SeparateKEPo-M51.0%46.8%75.0%51.1%
SeparateTopoPoison61.0%61.7%85.0%74.5%

Table 2: Attack effectiveness under consistent file organization against LightRAG on MuSiQue.

OrganizationMethodTASRC-TASRUASRC-UASR
OrderedGragPoison56.0%49.1%73.0%57.9%
OrderedKEPo-S53.0%47.4%67.0%49.1%
OrderedKEPo-M46.0%35.1%59.0%40.4%
OrderedTopoPoison75.0%75.4%86.0%77.2%
ShuffledGragPoison47.0%38.6%65.0%50.9%
ShuffledKEPo-S53.0%50.9%69.0%54.4%
ShuffledKEPo-M54.0%43.9%67.0%47.4%
ShuffledTopoPoison69.0%70.2%86.0%78.9%
SeparateGragPoison61.0%54.4%82.0%75.4%
SeparateKEPo-S74.0%68.4%84.0%73.7%
SeparateKEPo-M66.0%59.6%78.0%63.2%
SeparateTopoPoison85.0%82.5%92.0%86.0%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.