acceptodds
Under review as a conference paper at ICLR 2027

Backdoor in the Loop: Compromising Agentic Search via Malicious Retrievers

Abstract

Agentic retrieval-augmented generation (RAG) interleaves reasoning with repeated retrieval, giving the retriever influence over both the evidence an agent observes and its subsequent search decisions. We study retriever backdoors that exploit this feedback loop and repurpose weak backdoor purification to conceal their presence. An attacker supplies a compromised retriever checkpoint while leaving the search agent and deployment corpus unchanged. Without corpus write access, the attacker can still suppress useful evidence, persistently retrieve a selected existing document, or steer the agent toward prolonged search, inflating retrieval, context, and latency cost. To conceal these behaviors from detection, we propose leveraging a controlled inject-and-remove cycle: deliberately inject a weaker backdoor and then unlearn it. This process weakens detector-visible signatures and fools the backdoor detectors with an illusion of purification while preserving the malicious retrieval behavior. These findings expose a systematic vulnerability in RAG systems in which a weak defense becomes an attacker's concealment tool for a backdoored retriever, even when the underlying corpus remains trustworthy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Attack, Agentic Retrieval, Dense Retrieval, LLM Agents, AI Security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_g181OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work investigates backdoor attacks against retrievers in agentic Retrieval-Augmented Generation (RAG). The proposed attack assumes that an attacker can provide a malicious retriever checkpoint without modifying the search agent or knowledge database.

The authors propose three attacks: an untargeted attack that suppresses relevant evidence, a targeted attack that repeatedly retrieves a selected document, and a round-control attack that prolongs the agent's search process. The paper further proposes Decoy Overwrite–Unlearn (DOU), which injects and removes an auxiliary backdoor to reduce the detectability of the original backdoor while preserving its malicious behavior.

The authors evaluate the attacks on HotpotQA and TriviaQA using E5 and Contriever. The attacks substantially reduce answer accuracy, achieve nearly 100% persistent target retrieval, and increase search costs. DOU also reduces detection signals across several evaluated backdoor detectors.

Critical strengths
  • This work investigates retriever backdoors in agentic RAG, considering both answer manipulation and search efficiency. The three attack objectives demonstrate different security risks introduced by compromised retrievers.

  • The proposed DOU method is interesting because it exploits backdoor purification to conceal malicious behavior. The paper also evaluates multiple detection methods and reports the trade-off between concealment and attack effectiveness.

Critical weaknesses
  • The attack effectiveness evaluation is limited to HotpotQA and TriviaQA using a single agentic search framework. Furthermore, the attack relies on Agentic-R's design, which includes the original question in every retrieval round, allowing the trigger to persist. Its effectiveness in other agentic search frameworks remains unclear.

  • The backdoor detection evaluation is limited. The TriviaQA and MuSiQue detection sets contain only 8 and 30 naturally triggered questions, respectively. Although the experiments are repeated with different seeds, they reuse the same question sets. Furthermore, Neural Cleanse and PICCOLO share the same trigger-inversion procedure, limiting the diversity of evaluated detectors.

Questions for authors
  • Can the authors evaluate whether the attacks remain effective when the retriever receives only agent-generated queries? Since Agentic-R includes the original question in every retrieval round, the backdoor trigger remains present throughout the search process.

  • Can the authors provide additional end-to-end evaluations on benchmarks such as NQ, MuSiQue, or 2WikiMultiHopQA to demonstrate generalizability?

  • How reliable are the detection results given the limited number of naturally triggered questions? Can the authors provide evidence using larger evaluation sets and more diverse detection methods?

2 comments
AuthorsOpenReview ↗

Response to Reviewer g181 — Questions

Thank you for your careful review and constructive suggestions. We address your questions below.

Q1. Can the attacks remain effective when the retriever receives agent-generated queries?

Yes. The attack remains effective under DeFT-RAG, which receives the original question in the first round and reformulates subsequent queries. Using the same poisoned E5 on 1,000 paired HotpotQA questions, triggering increases target retrieval from 4.80% to 100% and reduces F1 from 50.00 to 33.22. Target hits occur in 64.44% of retrieval rounds, with an average of 4.586 rounds per triggered run.

We added these results to Section 4.2.2, “Transfer to DeFT-RAG,” and Table 5.

SettingEMF1Target HitHit/RoundTop-1 Hit
Backdoor-Clean37.7050.004.805.654.30
Backdoor-Trigger23.7033.22100.0064.44100.00

Target Hit and Top-1 Hit are sample-level metrics; Hit/Round uses executed retrieval rounds.

Q2. Can you provide additional end-to-end evaluations on benchmarks such as NQ, MuSiQue, or 2WikiMultiHopQA?

Yes. We added end-to-end evaluations on 2WikiMultiHopQA with two additional search agents, K2-Horizon-7B and Qwen3.5-27B. On 200 paired questions, without retraining the poisoned E5, triggering reduces F1 by 31.58 and 34.68 points, respectively, relative to Backdoor-Clean. The target ranks first in every executed triggered retrieval round.

We added these results to Section 4.2.2, “Transfer to additional search agents and a multi-hop benchmark,” and Table 4.

AgentSettingEMF1Target HitTarget@1/Round
K2-Horizon-7BOriginal Retriever46.5057.310.000.00
K2-Horizon-7BBackdoor-Clean44.0055.270.500.24
K2-Horizon-7BBackdoor-Trigger17.5023.6999.50100.00
Qwen3.5-27BOriginal Retriever57.5066.490.000.00
Qwen3.5-27BBackdoor-Clean55.0063.570.000.00
Qwen3.5-27BBackdoor-Trigger23.5028.8997.50100.00

Q3. How reliable are the detection results with few naturally triggered questions? Can you evaluate larger sets and more diverse detectors?

The reduction in detector separation persists on a larger evaluation set and across different detection mechanisms. We expanded detection to the full MuSiQue development set. DOU reduces separation for six of seven detectors, including representation- and perturbation-based methods; the expanded evaluation also includes ONION.

We added these results to Appendix D.3, “Concealment persists on the full MuSiQue development set,” and Table 14, complementing the main-text results in Table 9.

DetectorOriginal backdoorDOU80
Activation Clustering50.0010.89
Black-box TopK Instability44.009.86
ONION6.676.67
PICCOLO39.6820.72
RAP11.695.16
STRIP20.1415.56
Spectral Signatures50.0022.53

Values are detector separation Δ=meani∣AUROCi−50∣\Delta=\mathrm{mean}_i|\mathrm{AUROC}_i-50|; lower values indicate weaker separation. DOU80 uses 80 overwrite and 80 unlearn steps.

AuthorsOpenReview ↗

Response to Reviewer g181 — Weaknesses

W1. Limited benchmarks and dependence on Agentic-R’s inclusion of the original question.

The attack transfers beyond Agentic-R to DeFT-RAG and remains effective on the additional 2WikiMultiHopQA benchmark. DeFT-RAG reformulates subsequent queries after receiving the original question in the first round. Using the same poisoned E5 on 1,000 paired HotpotQA questions, triggering increases target retrieval from 4.80% to 100% and reduces F1 from 50.00 to 33.22. Target hits occur in 64.44% of retrieval rounds.

We added these results to Section 4.2.2, “Transfer to DeFT-RAG,” and Table 5.

SettingEMF1Target HitHit/RoundTop-1 Hit
Backdoor-Clean37.7050.004.805.654.30
Backdoor-Trigger23.7033.22100.0064.44100.00

Target Hit and Top-1 Hit are sample-level metrics; Hit/Round uses executed retrieval rounds.

We also evaluated 200 paired 2WikiMultiHopQA questions using K2-Horizon-7B and Qwen3.5-27B, without retraining the poisoned E5. Triggering reduces F1 by 31.58 and 34.68 points, respectively, relative to Backdoor-Clean. The target ranks first in every executed triggered retrieval round.

We added these results to Section 4.2.2, “Transfer to additional search agents and a multi-hop benchmark,” and Table 4.

AgentSettingEMF1Target HitTarget@1/Round
K2-Horizon-7BOriginal Retriever46.5057.310.000.00
K2-Horizon-7BBackdoor-Clean44.0055.270.500.24
K2-Horizon-7BBackdoor-Trigger17.5023.6999.50100.00
Qwen3.5-27BOriginal Retriever57.5066.490.000.00
Qwen3.5-27BBackdoor-Clean55.0063.570.000.00
Qwen3.5-27BBackdoor-Trigger23.5028.8997.50100.00

W2. Small detection sets and shared trigger inversion in NC/PICCOLO.

The concealment effect persists with more naturally triggered questions and across different detection mechanisms. We expanded detection to the full MuSiQue development set: 2,411 unique questions, including 73 naturally triggered positives. DOU reduces separation for six of seven evaluated detectors.

We added these results to Appendix D.3, “Concealment persists on the full MuSiQue development set,” and Table 14, complementing the main-text results in Table 9.

DetectorOriginal backdoorDOU80
Activation Clustering50.0010.89
Black-box TopK Instability44.009.86
ONION6.676.67
PICCOLO39.6820.72
RAP11.695.16
STRIP20.1415.56
Spectral Signatures50.0022.53

Values are detector separation Δ=meani∣AUROCi−50∣\Delta=\mathrm{mean}_i|\mathrm{AUROC}_i-50|; lower values indicate weaker separation. DOU80 uses 80 overwrite and 80 unlearn steps.

NC and PICCOLO originate from distinct detection methods, although both use trigger inversion and our retrieval adaptations share this component. Our evaluation also includes Activation Clustering, Spectral Signatures, Black-box TopK Instability, RAP, STRIP, and ONION. Thus, the evidence for concealment extends across multiple detection mechanisms rather than relying solely on NC and PICCOLO.

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

The paper presents a set of attacks against agentic RAG systems that operate over multiple rounds of interaction. In each round, the agent decides whether to refine its query or issue an additional one to retrieve more context, or to stop and answer the original query using all the context collected so far.

The paper studies attacks that inject a backdoor into the retriever model through fine-tuning, while leaving the agent and the document corpus intact. In this setting, the paper explores three attacks: (1) an untargeted attack, in which a query containing a trigger phrase causes the retriever to prioritize less similar documents; (2) a targeted attack, in which the retriever prioritizes a specific attacker-chosen document; and (3) a prolonged search attack, which aims to increase the number of rounds before the agent stops and generates a final answer.

To evade backdoor detection defenses, the paper proposes a concealment approach that embeds an additional decoy backdoor and then unlearns it. The resulting model is less likely to be flagged by defenses, while retaining the original backdoor behavior.

The attacks and the concealment approach are evaluated on two retriever models and two datasets, and three trigger types.

Critical strengths
  • Novel setting. The paper investigates a novel class of attacks against agentic multi-round RAG systems, whereas most prior work in this area focuses on the traditional single-round RAG setting.

  • Prolonged search attack. The prolonged attack is a particularly interesting contribution. Denial-of-service and degradation-of-service attacks are underexplored and are highly interesting. In addition, this attack is tailored directly to the round-based nature of agentic RAG systems, unlike the other two attacks can also be applied to traditional RAG settings.

  • Round-level retrieval metrics. I appreciated the round-level retrieval metrics, which report per-round results in addition to the final response. These results are impressive and show that the target document is retrieved in nearly every round.

  • Concealment. I found the proposed concealment approach to be novel and interesting. Using a defense technique to circumvent defenses is a clever idea, and the results are promising.

  • Extensive evaluation of DOU. The evaluation of DOU is extensive, covering many defenses that target different properties of backdoored models.

Critical weaknesses
  • Clarity. The paper is difficult to read. While the high-level idea is easy to understand, the inner workings of the attacks and of DOU are hard to follow due to confusing and inconsistent notation. I also felt like the paper assumes substantial background from the reader. Higher-level descriptions of the purpose of each part of the evaluation would also help. For example, the purpose of Section 4.3 was initially unclear, and it took some time to understand that it motivates DOU by showing that purification approaches do not remove the backdoor.

  • Missing details and a single agent model. Some hyperparameters and design choices are neither reported nor ablated, such as tau in Section 2.2 and the judge model in Section 2.3. In addition, the model used as the RAG agent is mentioned only in the appendix, and only one agent model is evaluated. Since the agent determines which queries to issue next and what to include in the final response, this choice could substantially affect the results, and additional agent models should be evaluated.

  • Threat model. The threat model is stronger than presented. The paper argues that its threat model is highly realistic because it does not require modifying the agent or the corpus. However, the attack relies on access to both in order to fine-tune the retriever and inject the backdoor. This is a strong assumption, and no experiment shows that it can be relaxed, for example by training without access to the actual corpus, or by evaluating the prolonged search attack with a different agent than the one used during training.

  • Metrics. The chosen metrics are not sufficient to show that the attacker’s goal is achieved. They mainly show that the answer generated under the trigger differs from the clean answer, but not which answer is obtained instead. Similarly, the retrieval metrics show that the attacker-chosen document is retrieved, even at every round, but not that this causes the agent to adopt the content of that document in its final response. This could be addressed by adding a semantic metric (in contrast to lexical metrics such as EM and F1) that compares the response to both the clean response and the attacker’s target. Some examples are provided in the appendix, but they should be highlighted in the main paper and quantified.

  • Limited analysis of the concealment method. The paper presents the concealment method as one of its main contributions but does not sufficiently discuss why it works. While Section 4.3 shows that purification methods do not remove the backdoor, it remains unclear why the resulting model is less likely to reveal the backdoor behavior. In addition, the paper evaluates many defenses without explaining in the main text what each one measures (this appears only in the appendix), which makes the reported results difficult to interpret in my opinion. A more in-depth discussion of this phenomenon would strengthen the paper, since it is its most novel aspect. It would also be valuable to understand how the choice of decoy affects the method, for instance whether the decoy must be sufficiently different from the original trigger for the approach to work.

Questions for authors
  • How do the attacks perform when the attacker does not have access to the actual clean corpus and RAG agent during training?

  • What are the results when other models are used as the RAG agent?

  • Could you clarify why DOU makes the backdoor harder to detect while preserving nearly all of its effectiveness? Please also elaborate on how the choice of decoy affects the results.

  • Section 4.3 shows that other purification methods (e.g., fine-tuning on additional clean data) also fail to remove the backdoor. Could these methods also be used for concealment? This relates to the previous question: what specifically about the decoy learning-and-unlearning approach makes it effective?

2 comments
AuthorsOpenReview ↗

Response to Reviewer zBYT — Questions

We appreciate the opportunity to clarify these points and provide additional experimental evidence.

Q1. How do the attacks perform without access to the deployed corpus and RAG agent during training?

The untargeted attack remains effective without access to the deployment corpus or the deployed RAG agent during training. Training solely on MS MARCO gives:

DatasetClean F1Triggered F1
HotpotQA56.0533.36
TriviaQA71.4850.96

Targeted retrieval requires knowledge of the selected existing document, while round-control training uses offline trajectories or a reference agent to construct its targets (Sections 2.2 and 3.3). The results above demonstrate corpus transfer for the untargeted attack.

Q2. What are the results when other models are used as the RAG agent?

The attack remains effective with K2-Horizon-7B and Qwen3.5-27B without retraining the poisoned retriever:

We added these results to Section 4.2.2, “Transfer to additional search agents and a multi-hop benchmark,” and Table 4.

AgentInputEMF1Target Hit
K2-Horizon-7BClean44.0055.270.50
K2-Horizon-7BTriggered17.5023.6999.50
Qwen3.5-27BClean55.0063.570.00
Qwen3.5-27BTriggered23.5028.8997.50

The attack also transfers to DeFT-RAG, which reformulates subsequent queries. Target hit increases from 4.80% to 100%, while F1 falls from 50.00 to 33.22 (Section 4.2.2, “Transfer to DeFT-RAG,” and Table 5).

Q3. Why does DOU conceal the backdoor while preserving effectiveness, and how does decoy choice affect the results?

DOU changes detector-visible signatures while preserving the relative document rankings needed by the attack. Targeted retrieval remains effective when the target stays above competing documents at the top-K boundary; untargeted retrieval remains effective when useful evidence is displaced. Neither requires query representations, clustering structure, or perturbation responses to remain unchanged.

Starting from the compromised retriever, overwriting introduces a separate auxiliary trigger and trains its queries toward mismatched documents. Unlearning then restores those queries' correct document associations. Restoring an association does not reverse the parameter updates: the second stage optimizes a different objective, including parameter regularization, from the checkpoint produced by overwriting. Clean supervision constrains utility loss, while limited updates allow the original malicious rankings to survive.

The comparison with overwrite-only processing supports the importance of the subsequent unlearning stage:

ProcessingNC: Mean |AUROC − 50| ↓PICCOLO: Mean |AUROC − 50| ↓
Overwrite only35.0536.36
DOU (80/80)19.4021.64

AUROC is expressed as a percentage; lower values indicate detection closer to chance.

Decoy choice determines the auxiliary associations through which these updates occur. We use a distinct trigger, bpooverwrite, with an all-to-all permutation mapping each auxiliary query to another query's positive document. This distributes associations across documents, whereas an all-to-one decoy concentrates them around a shared target. A common target can retain a detectable representation direction. Consistently, the original targeted backdoor is harder to conceal: AC's mean |AUROC − 50| decreases only from 35.29 to 30.61 (Section 4.5 and Table 9).

The mechanism and decoy construction are described in Section 3.4 and Appendix A.2. These experiments support the selected construction; sensitivity to alternative decoy tokens and assignments has not been systematically measured.

Q4. Can other purification methods also conceal the backdoor, and what is specific to decoy learning–unlearning?

Clean fine-tuning provides substantially less concealment than DOU, although both preserve the attack. With the same total training budget—160 CleanFT steps versus 80 overwrite plus 80 unlearn steps—we obtain:

MethodACNCPICCOLOSSTriggered F1
CleanFT35.9839.3542.8142.2547.57
DOU (80/80)23.8319.4021.6429.2646.62

Detector columns report mean |AUROC − 50| on MuSiQue; triggered F1 is on TriviaQA.

Both retain similar answer degradation, but DOU brings detection substantially closer to chance. Its distinguishing feature is the temporary decoy association followed by its removal; ordinary clean supervision does not produce the same effect. Section 4.3 and Table 7 establish persistence under purification; this additional comparison tests concealment.

AuthorsOpenReview ↗

Response to Reviewer zBYT — Weaknesses

We appreciate your suggestions for improving the clarity and evaluation of our work.

W1. Clarity.

We believe the paper is reasonably clear overall, while acknowledging that some explanations can be improved. Both reviewers QM8w and g181 rated the presentation 3 (“Clear and easy to follow”). Reviewer g181 also explicitly recognized that DOU “exploits backdoor purification to conceal malicious behavior” and that the paper reports the trade-off between concealment and attack effectiveness.

We nevertheless appreciate your feedback and will clarify the attack steps, unify notation, and state each evaluation's purpose upfront. Section 4.3 tests whether purification removes the original backdoor and motivates the subsequent DOU analysis. Please see Q3 for the mechanism and role of decoy choice.

W2. Missing details and a single agent model.

Please see Q2 for the additional evaluations with K2-Horizon-7B, Qwen3.5-27B, and DeFT-RAG.

W3. Threat model.

Please see Q1 for the experiment without access to the deployment corpus or deployed agent during training, and the access requirements of the different attack objectives.

W4. Metrics.

The untargeted attack predominantly turns previously correct answers into wrong answers, rather than merely causing the agent to return no answer. On HotpotQA, 54.4–58.3% of previously correct answers become wrong answers, compared with 6.5–9.9% becoming missing answers. On TriviaQA, these proportions are 35.0–37.7% and 3.1–4.0%, respectively.

We added these results to Section 4.2.1, “Wrong answers dominate triggered failures,” and Table 2.

DatasetTriggerPreviously correctWrong answerNo answer
HotpotQAauthor425231/425 (54.4%)42/425 (9.9%)
HotpotQAhow432249/432 (57.6%)32/432 (7.4%)
HotpotQA##429250/429 (58.3%)28/429 (6.5%)
TriviaQAauthor647244/647 (37.7%)26/647 (4.0%)
TriviaQAhow643225/643 (35.0%)20/643 (3.1%)
TriviaQA##639241/639 (37.7%)21/639 (3.3%)

Semantic Answer Similarity (SAS) further shows substantial changes in final responses. Clean-to-trigger similarity falls to 31.03–32.72 on HotpotQA and 48.41–54.28 on TriviaQA, compared with Original self-similarity of 92.76 and 94.40.

We added these results to Appendix C, “Semantic answer similarity,” and Table 11.

DatasetSettingSAS ↑
HotpotQAOriginal92.76
HotpotQAauthor31.03
HotpotQAHow32.58
HotpotQA##32.72
TriviaQAOriginal94.40
TriviaQAauthor48.41
TriviaQAHow53.12
TriviaQA##54.28

Original compares each answer with itself; trigger rows compare clean and triggered answers.

W5. Limited analysis of the concealment method.

Please see Q3 for the mechanism, overwrite-only comparison, and decoy construction, and Q4 for the comparison with clean fine-tuning.

The detectors examine different signals: NC and PICCOLO use trigger inversion, AC and SS analyze representations, and the perturbation-based methods probe embedding or retrieval responses.

Reviewer_QM8wOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies retriever-only backdoor attacks on agentic RAG systems, without modifying the search agent or the corpus. The authors propose three types of attacks that aim to lower the answer quality, repeatedly retrieve a specific targeted document, or increase the number of search rounds. They also introduce DOU, an approach that repurposes backdoor purification techniques to make the attacks harder to detect. The authors conduct experiments on two QA datasets, two retrievers, and a single search agent.

Critical strengths
  • The security concern (backdoor attack) of the RAG system is important in the current agentic search systems, while there is a lot of prior work in this area.

  • I think the proposed DOU is an interesting approach that repurposes existing backdoor purification techniques to conceal malicious behavior.

Critical weaknesses
  • A major concern is the limited practical scope of the proposed threat model in a real-world setting. Specifically, non-neural retrieval methods (e.g., BM25) are (still) widely used in practice due to their effectiveness and efficiency (not requiring any training or calling API), whereas the proposed attacks assume only neural dense retriever checkpoints.

  • Related to the above concern, simple alternatives (e.g., BM25 or hybrid retrieval) could potentially mitigate such attacks, which (again) makes me question the practical significance of the proposed retriever-only attacks, which the authors argue and highlight as a key contribution.

  • Also, the novelty of the proposed attacks in the agentic setting is limited. Retriever backdoor attacks have already been studied in prior work (as the authors also acknowledge in the paper), and it is unclear what is fundamentally new here (beyond repeatedly activating the same backdoor across retrieval rounds). In particular, since the original question (including the trigger) is included in every retrieval input, the attack may simply be reactivated at each round, rather than leveraging any unique property of agentic search.

  • I am also concerned that the experiments and evaluations do not sufficiently reflect “real” realistic agentic settings. The experiments are conducted on two conventional QA benchmarks (HotpotQA and TriviaQA), which were not specifically designed for agentic tasks (which were widely used and constructed before the era of “LLM agents”), rather than more recent and complex agentic workflows. Also, the observed search trajectories are relatively short, so the proposed attacks may not remain effective in more complex, long-horizon agentic settings, which would be more challenging.

  • The evaluation scope is also limited. While the paper focuses on retriever-side attacks, evaluation on only two dense retrievers seems largely insufficient to show the generalizability of the proposed argument. Also, the experiments are based on only a single search agent, which I think is more problematic, as I believe that stronger (and more recent) LLMs may be (much) better at identifying and filtering out irrelevant retrieved documents, potentially mitigating the impact of such attacks.

Questions for authors

See above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer QM8w — Weaknesses (Part 1/2)

Thank you for your careful review. We address the practical relevance of our threat model and the role of hybrid retrieval below.

W1. Practical relevance of the dense-retriever threat model.

E5-based dense retrieval is used in multiple recent agentic search systems, making the integrity of this component a relevant security concern. Agentic-R [1] initializes its retriever from E5-base-v2 and trains it for evolving agent queries (Appendix B). AutoSearch [2] uses E5 within its adaptive multi-step search framework (Section 5.1). ESearch [3], published at ICLR 2026, also uses E5 as the retriever in its search-augmented reasoning framework. All three are published 2026 studies using the same dense-retriever family as our E5 experiments.

These works illustrate the continued adoption of neural retrievers for complex agentic tasks, where semantic matching beyond lexical overlap supports evolving search queries. In particular, Agentic-R improves retrieval by incorporating both local relevance and downstream answer correctness into retriever training.

Our threat model concerns systems that deploy an externally supplied dense retriever. A compromised checkpoint can change the evidence returned to the agent and thereby affect subsequent queries, stopping decisions, and final answers. Our experiments cover E5 and Contriever while keeping the agent and corpus unchanged.

Our additional E5–BM25 experiment further shows that incorporating lexical retrieval does not eliminate the vulnerability of the compromised dense component, as detailed in W2.

[1] Liu et al. (2026). Agentic-R: Learning to Retrieve for Agentic Search. Findings of ACL 2026. https://aclanthology.org/2026.findings-acl.785/

[2] Sun et al. (2026). AutoSearch: Adaptive Search Depth for Efficient Agentic RAG via Reinforcement Learning. Findings of ACL 2026. https://aclanthology.org/2026.findings-acl.1399/

[3] An et al. (2026). Erase to Improve: Erasable Reinforcement Learning for Search-Augmented LLMs. ICLR 2026. https://openreview.net/forum?id=UpZHjUtcUM

W2. Whether BM25 or hybrid retrieval mitigates the attack.

Adding BM25 does not eliminate the attack in our hybrid evaluation. We combined the same backdoored E5 retriever with BM25 and evaluated 1,000 paired HotpotQA development questions:

InputEMF1AccPrecisionRecall
Clean39.9050.9442.3054.2350.09
Triggered29.1037.3530.6039.5936.92

Triggering reduces EM by 10.80 points and F1 by 13.59 points, despite the presence of BM25. Thus, a hybrid system can still inherit the vulnerability of its compromised dense component.

A BM25-only system falls outside our retriever-checkpoint threat model because it does not use the learned dense retriever being attacked. The hybrid result addresses the more direct question of whether adding lexical retrieval protects a system that continues to use that compromised component.

AuthorsOpenReview ↗

Response to Reviewer QM8w — Weaknesses (Part 2/2)

We address the distinction from conventional RAG attacks and the breadth of our evaluation below.

W3. What is specific to attacking agentic search?

The distinction is that corrupted retrieval can affect the agent's subsequent search behavior, beyond the evidence used for a single answer. In agentic search, retrieved passages influence later queries and stopping decisions. Our three objectives therefore examine evidence degradation, persistent target retrieval across rounds, and control of search duration.

Retaining the original question in successive retrieval inputs is used across different 2026 retrieval frameworks, beyond Agentic-R. ThinkRetrieve [4] jointly encodes the original question and the intermediate answer at each retrieval step (Eq. 5 and Algorithm 1). SARDI [5] concatenates the original question with the current intermediate response for retrieval at each denoising iteration (Eq. 4). RVR [6] augments the original query with previously verified documents for subsequent retrieval rounds. These examples show that retaining the original question is a design shared by multiple existing architectures, giving our setting relevance across a broad range of iterative retrieval systems.

The attack also remains effective when the agent reformulates its queries in subsequent rounds. DeFT-RAG generates successive sub-questions. With the same poisoned retriever, clean and triggered inputs yield target-hit rates of 4.80% and 100.00%, respectively, while answer F1 decreases from 50.00 to 33.22 on HotpotQA. Target hit counts a trajectory as successful if the selected document is retrieved in at least one search round; it does not require retrieval in every round. These results show that query reformulation does not eliminate the observed target-retrieval effect or answer degradation.

We added these results to Section 4.2.2, “Transfer to DeFT-RAG,” and Table 5.

InputEMF1Target Hit
Clean37.7050.004.80
Triggered23.7033.22100.00

Round control also exposes a consequence specific to an iterative workflow: prolonged search consumes additional resources. Section 4.2.3 and Table 6 report increases of approximately 78–88% in retrieval rounds, 81–87% in context tokens, and 80–126% in latency.

DOU adds a further security finding: processing intended to purify a model can instead reduce backdoor detectability while preserving its effects. Sections 4.3–4.5 distinguish persistence under purification from concealment after decoy overwriting and unlearning.

W4. Applicability to different search architectures and longer trajectories.

The attack transfers to a different search architecture, with target retrieval observed in trajectories extending to nine retrieval rounds. Please see W3 above for the DeFT-RAG results, included in Section 4.2.2, “Transfer to DeFT-RAG,” and Table 5.

The triggered runs average 4.59 retrieval rounds, with 40 runs reaching nine rounds. The target is retrieved at least once in every evaluated trajectory.

W5. Generalization beyond the original agent and benchmarks.

The attack remains effective with K2-Horizon-7B and Qwen3.5-27B on 2WikiMultiHopQA, without retraining the poisoned retriever.

We added these results to Section 4.2.2, “Transfer to additional search agents and a multi-hop benchmark,” and Table 4.

AgentInputEMF1Target Hit
K2-Horizon-7BClean44.0055.270.50
K2-Horizon-7BTriggered17.5023.6999.50
Qwen3.5-27BClean55.0063.570.00
Qwen3.5-27BTriggered23.5028.8997.50

Both agents exhibit substantial answer degradation and high target-hit rates. Together with the DeFT-RAG results in W3–W4 and the hybrid-retrieval results in W2, these experiments extend the evidence across agent models, a different search architecture, an additional multi-hop benchmark, and a lexical–dense retrieval setting.

[4] Singh et al. (2026). ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling. Findings of EMNLP 2026. https://aclanthology.org/2026.findings-emnlp.1703/

[5] Jünger et al. (2026). Self-Augmenting Retrieval for Diffusion Language Models. ICML 2026. https://proceedings.mlr.press/v306/junger26a.html

[6] Qian et al. (2026). RVR: Retrieve-Verify-Retrieve for Comprehensive Question Answering. arXiv:2602.18425. https://arxiv.org/abs/2602.18425

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.