acceptodds
Under review as a conference paper at ICLR 2027

Cross-Path Poisoning Attacks on Structured RAG

Abstract

Retrieval-Augmented Generation (RAG) systems increasingly combine unstructured document retrieval with structured database access, allowing language models to synthesize vector-retrieved evidence and SQL results within a single pipeline. We show that this integration introduces a previously underexplored cross-path poisoning surface: relational content modified through legitimate database operations can be serialized, chunked, and indexed as retrievable text, allowing a single structured data source to influence both reasoning paths. To the best of our knowledge, this is the first systematic study of poisoning attacks that originate from relational data and propagate across both the structured SQL and vector-retrieval paths of a hybrid RAG system. We study three realizations of this threat—Update, Insert, and a higher-privilege Vector extension—while requiring the correct top-ranked SQL result to remain unchanged. Across Qwen, Llama, and Mistral, the frozen attack configurations induce incorrect answers in all 50 fixed-query runs for every attack–model pair and retain a 93.4% aggregate error rate over 40 paraphrased query variants. Our analysis shows that the vulnerability emerges from the composition of data serialization, relevance-based reranking, and final evidence synthesis: misleading database content can be promoted as highly relevant retrieved evidence and subsequently override independently computed SQL results. We further study retrieval-side anomaly signals, finding strong separation for vector-level poisoning but a substantial detection–false-positive trade-off for text-based attacks. Beyond the specific system evaluated here, our threat model applies to hybrid RAG architectures in which structured records are reused through both deterministic database queries and semantic retrieval, establishing cross-path consistency as a security requirement for this broader class of systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-Augmented Generation, Structured RAG, RAG Security, Data Poisoning, Cross-Path Poisoning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_kifpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies cross-path poisoning in hybrid RAG systems where relational data are used both for SQL-based reasoning and, after serialization, for vector retrieval. It introduces UPDATE, INSERT, and VECTOR attacks that manipulate relational content so that poisoned retrieval evidence can override correct structured SQL results during final answer generation.

Critical strengths
  1. The UPDATE and INSERT attacks explore a relatively novel attack surface by poisoning relational data and allowing the system’s normal ingestion pipeline to propagate the malicious content into the retrieval path, rather than directly modifying the retrieval corpus.
  2. The evaluation covers three attack variants and examines multiple dimensions, including different LLMs, query paraphrases, injection volume, and retrieval-side detection signals.
Critical weaknesses
  1. The practical threat model is insufficiently motivated. The paper assumes legitimate write access to selected relational fields, but does not clearly instantiate realistic attacker roles or explain how commonly such attacker-controlled fields would subsequently be ingested into the RAG corpus.
  2. The paper lacks direct experimental comparison with existing RAG poisoning attacks, making it difficult to assess how much additional attack capability is provided specifically by the proposed cross-path setting.
  3. There is a severe internal inconsistency in the main effectiveness result. The abstract and Section 4.1 state that every attack–model pair causes an incorrect answer in all 50 fixed-query runs. However, the table immediately described as reporting those incorrect answers contains values such as 14/50 for UPDATE–Qwen, 39/50 for UPDATE–Llama, 27/50 for INSERT–Qwen, and 26/50 for VECTOR–Mistral; only some VECTOR rows are 50/50. One row even shows UPDATE–Mistral with fewer errors under ATTACK (26) than NEUTRAL (40). As written, the headline claim cannot be reconciled with the primary table. This is a critical issue because attack effectiveness is the central empirical claim of the paper.
Questions for authors

Resolve the issues raised in the "weaknesses"

Reviewer_1vooOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates cross-path poisoning attacks against structured RAG systems that combine SQL queries with vector retrieval. The proposed attacks modify relational database content, which is subsequently serialized, embedded, and retrieved as textual evidence. The authors aim to manipulate the final answer while preserving the correct SQL results. The paper introduces three attacks: UPDATE, which modifies existing text cells; INSERT, which adds malicious rows containing structured formatting; and VECTOR, which directly injects retrieval vectors and associated text under a higher-privilege threat model. The authors evaluate these attacks across Qwen, Llama, and Mistral using repeated queries and paraphrased variants. They also investigate a retrieval-side anomaly detector based on retrieval and reranking behavior.

Critical strengths
  • This work investigates an interesting security problem in structured RAG systems, where poisoned relational content can influence both SQL-based reasoning and vector retrieval.

  • The paper proposes three attack strategies with different attacker capabilities and investigates how poisoning propagates through serialization, retrieval, reranking, and final answer generation.

Critical weaknesses
  • The experimental evaluation is insufficient. The authors evaluate only one structured database scenario with seven small tables. Furthermore, the experiments mainly rely on repeated executions of a fixed query and paraphrases of the same underlying question. Although the authors acknowledge this limitation, the generalizability of the attacks remains unclear.

  • The experimental reporting raises substantial concerns. Although the fixed-query experiments report 50/50 incorrect answers across all attack–model combinations, the matched-control experiments show inconsistent attack effectiveness. For example, UPDATE against Mistral produces fewer errors than the neutral control (26/50 vs. 40/50), which the authors acknowledge. Furthermore, the supplementary defense evaluation reports 0/10 attack success for UPDATE without additional defenses. The differences in evaluation settings and success metrics are not sufficiently explained.

Questions for authors
  • Can the authors provide additional evaluations across different independent queries, database schemas, and structured RAG pipelines?

  • Can the authors clarify the differences between the fixed-query, matched-control, and defense-ablation experiments, including their attack configurations and success metrics? In particular, why does UPDATE report 50/50 errors in the fixed-query evaluation but 0/10 attack success without additional defenses?

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

This paper studies poisoning attacks on hybrid RAG systems that combine SQL results with retrieved text. The key observation is that database fields can become misleading evidence after serialization and indexing, causing the final answer to contradict a correct SQL result. The authors develop UPDATE and INSERT attacks through database writes, plus a higher-privilege VECTOR variant. Experiments cover three model families, query paraphrases, injection volume, and several detection and defense settings.

Critical strengths
  • The database entry point is worth studying. In particular, UPDATE and INSERT show how writable text fields can influence answer generation through normal ingestion. Keeping the correct SQL winner unchanged helps isolate the failure at the retrieval and synthesis stages.

  • The matched neutral controls and injection-volume experiments are useful. They help distinguish the effect of a misleading claim from that of an arbitrary database modification, and show that effectiveness can depend strongly on how much content is injected.

Critical weaknesses
  • The headline effectiveness is not reconciled with the other experiments. Appendix Table 2 reports 50/50 errors for every attack–model pair, while main-paper Table 1 reports substantially lower counts for several pairs, including 14/50 for UPDATE on Qwen. Appendix Table 6 reports 0/10 for UPDATE in the undefended system. These experiments may use different settings, but those differences are not explained well enough to interpret the results. For UPDATE on Mistral, the matched neutral condition even causes more errors than the attack. In addition, the meaning of “50 independent runs” is unclear given the default temperature and seed of zero. I cannot tell how much weight to place on the headline success rates.

  • The evaluation is too narrow to support the broader conclusions. The main robustness experiment uses paraphrases of the same underlying question in one database scenario. This tests sensitivity to wording, rather than generalization across tasks. The answer is already determined by SQL, and enforcing agreement with SQL blocks all tested attacks in Appendix Table 6. The paper therefore leaves open the more consequential case where answering genuinely requires both textual and structured evidence. It also attributes success to serialization and reranking without reporting the defined delivery metrics or ablations that isolate their contributions.

  • The attacker’s knowledge and observation capabilities are not clearly specified. The paper describes database write permissions, but attack construction appears to require additional access to the RAG pipeline. INSERT selection uses top-3 coverage and reranking scores; VECTOR uses actual system-generated subqueries and the same embedding model as the retriever. It is unclear how the attacker obtains these signals, what feedback is available during optimization, and whether the synthesis prompt is known. VECTOR additionally grants control over stored retrieval vectors. These are materially different capabilities, and the paper does not establish a realistic deployment scenario for their combination. This makes the practical significance of the threat model difficult to assess.

  • The main text does not provide enough experimental evidence for its claims. Its only results table presents the matched-control analysis, although Section 4.1 refers to it as support for the fixed-query and paraphrase results. The headline effectiveness, injection-volume results, and defense ablations are instead placed in the supplement. Consequently, the main text asserts several central findings without presenting the corresponding evidence or enough detail to assess them. Even with the supplement, important information remains missing, including clearly identified final attack configurations and the delivery and targeted-success results introduced in Section 2.2. This is a substantive reporting gap, not simply a preference about where tables should appear.

Questions for authors
  • What differs between main-paper Table 1, Appendix Table 2, and Appendix Table 6 in models, queries, payloads, injection volume, and decoding settings? What varies across the 50 “independent” fixed-query runs?

  • For each attack, what can the attacker observe beyond database contents and final answers? What realistic deployment grants these capabilities?

  • How many distinct underlying tasks are used for development and evaluation, excluding paraphrases? Do any require textual evidence beyond what SQL already provides?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.