acceptodds
Under review as a conference paper at ICLR 2027

CyclicRA: Inducing Retrieval Loops for Resource Exhaustion in Agentic RAG

Abstract

Agentic RAG enables large language model (LLM) agents to iteratively coordinate reasoning, retrieval, and action based on evolving execution contexts, where retrieved evidence continuously affects subsequent execution. This iterative dependence exposes a new attack surface: poisoned documents can steer the agent toward unnecessary reasoning and retrieval, substantially increasing resource consumption. However, existing Denial-of-Efficiency attacks largely focus on direct manipulation of agent inputs, leaving the resource-exhaustion risks arising from the retrieval process itself largely unexplored. To systematically exploit this attack surface, we propose CyclicRA, a corpus-poisoning attack that steers the agent's retrieval trajectory by inducing self-sustaining retrieval loops. CyclicRA generates poisoned documents that provide seemingly useful evidence while introducing unresolved information gaps, encouraging the agent to issue additional retrieval queries. It further guides the agent through a sequence of intermediate queries, aiming to form a retrieval loop in which subsequent queries return to a previously issued query. This loop repeatedly extends the agent's reasoning and retrieval process beyond what is required for task completion. Experiments across five models on both single-hop and multi-hop question-answering benchmarks show that CyclicRA induces an average of 11.63 additional retrieval steps over clean execution and consumes 1.82x as many tokens as SOTA baseline attacks on average, while remaining effective under different defense measures.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agentic RAG, Corpus Poisoning, Denial-of-Efficiency, Retrieval Loops

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dhHKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a new attacks against Agentic RAG. The attack makes the agent to believe it needs to retrieve more information from the RAG, in the end resulting in a loop. The authors show how to construct the series of documents that will poison the corpus, based on the queries the attacks wants to induce on the agent. They demonstrate the attack on several models and they also compare it with two simple defenses.

Critical strengths

The problem setting is realistic from the point of view of attacker capabilities and knowledge. They demonstrate that creating the documents to induce the loop is indeed necessary, vs just having a random sequence of documents.

Critical weaknesses

There are system level defenses that can mitigate the attack or raise the bar for the attacker that are not explored such us preventing RAG poisoning in the first place by using security policies that control the documents accepted in the RAG. For web based documents indeed that might be a bit more difficult to control but it will also be more difficult for the attacker.

I did not see an evaluation of the overhead needed to generate the poisoned documents and the effectiveness of the attack based on the length of the loop.

Questions for authors

How long does it take to generate the poisoned documents based on the length of the loop?

Does the length of the loop matter for attack effectiveness?

3 comments
AuthorsOpenReview ↗

Response to W1

[W1] Could broader system-level defenses, including document admission policies, mitigate the attack or raise the attacker’s barrier?

System-level defenses can raise the attacker's barrier but do not necessarily eliminate corpus-poisoning risks. Our threat model assumes that the attacker can insert a small number of poisoned documents into the retrieval corpus, for example through a contributor account in a shared enterprise knowledge base. This setting is consistent with PoisonedRAG [1], which explicitly considers insider insertion into enterprise knowledge bases and poisoning through web sources. CorruptRAG [2] further demonstrates that effective RAG poisoning can require only a single poisoned text per target query. These studies support limited corpus poisoning as a valid threat model.

We have already evaluated two distinct defenses, Perplexity Filtering and Keyword-Density Filtering, to filter and control the documents actually read by Agentic RAG, and verified that CyclicRA remains effective under these defenses (Section 5.2).

To further strengthen our evaluation, we additionally consider a more sophisticated, behavior-aware system-level defense, LoopGuard, which monitors agent execution to detect repetitive and cyclic behavior. We evaluate LoopGuard on 200 SQuAD attack trajectories across four models. The results are presented below:

Table.R1: LoopGuard warning and interception rates on 200 SQuAD attack trajectories (%).

ModelMethodWarning rate (%)Interception rate (%)
Qwen3.5-4BOverThink7261
STA4210
CyclicRA3713
Qwen3.5-9BOverThink7854
STA5239
CyclicRA4721
Phi4-14BOverThink5942
STA4717
CyclicRA4122
Gemma3-12BOverThink5639
STA2611
CyclicRA226

CyclicRA achieves the lowest warning rate across all evaluated models and the lowest interception rate on two, suggesting that its nonconsecutive query revisits are harder for LoopGuard to detect under the current configuration.

[1] Zou, W., et al. (2025). PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models. Proceedings of the 34th USENIX Security Symposium, 3827–3844.

[2] Zhang, B., et al. (2026). Practical Poisoning Attacks against Retrieval-Augmented Generation. arXiv preprint arXiv:2504.03957.

AuthorsOpenReview ↗

Response to W2, Q1 adn Q2

[W2, Q1] What is the overhead of generating poisoned documents, and how does it vary with cycle length?

Poisoned-document generation takes minutes per task and has low estimated monetary costs, both increasing with cycle length. We measure offline attack-construction costs on the same 100 SQuAD questions with cycle lengths (k ∈ {3,5,7}). We report mean wall-clock time and total auxiliary-model tokens per task. API-equivalent costs are estimated using Qwen2.5-Coder-7B-Instruct at $0.05 per million input/output tokens.

Table.R2: Offline attack-construction costs across cycle lengths on 100 SQuAD tasks.

Cycle length kMean preparation time (s)Auxiliary-model tokens per taskEstimated cost per task (USD)
3117.1424524.2$0.00123
5157.5042807.4$0.00214
7231.6055796.6$0.00279

These minute-scale preparation times are practical for offline attacks, while the estimated monetary costs are negligible, supporting the practical feasibility of attack construction.

[W2, Q2] Does the length of the loop matter for attack effectiveness?

Cycle length affects attack effectiveness, but longer cycles do not consistently yield greater resource amplification. We compare cycle lengths k ∈ {3,5,7} on the same 200 SQuAD questions, keeping all other settings fixed.

Table.R3: Resource amplification by cycle length on the same 200 SQuAD tasks.

Victim modelCycle length kTCRΔSTEPSΔCAP HITS (pp)
Qwen3.5-4B332.399.148
535.4410.2010
758.5414.0318
Qwen3.5-9B369.105.3127
597.429.1929
788.797.6014

Cycle length introduces a trade-off between resource consumption and attack success. Longer cycles can incur greater retrieval overhead per completed cycle, but require more successful query transitions, increasing the risk of breaking the cycle. Consequently, longer cycles do not necessarily lead to greater overall resource amplification.

I have read the answer and will raise my score.

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

CyclicRA poisons documents to prolong retrieval in Agentic RAG. It constructs cyclic query dependencies intended to make the agent revisit earlier searches and consume additional resources.

Critical strengths
  1. The paper studies resource consumption as a consequence of corpus poisoning. This is a relevant concern for agents that repeatedly retrieve external information.
  2. The two-stage construction is understandable. The supplied algorithms explain how the intended query cycle guides document generation.
  3. The results show increased retrieval activity across several model families. The no-loop comparison provides initial evidence that document coordination matters.
Critical weaknesses
  1. The no-loop baseline also removes cross-document coordination, so it does not isolate cycle closure. Could the authors change only the final redirection to create an open-chain control? Please report the fraction of runs that actually complete the intended cycle.
  2. The NLI model is trained to judge entailment between statements. Why should its score on two questions predict the agent’s next search action?
  3. Equation (3) rewards the sum of dependency scores. This can favor a cycle containing one weak transition that prevents repeated traversal. Please compare against an objective that maximizes the weakest edge under the same search budget.
  4. The token ratio includes repeated processing of accumulated context. Please compare methods using equally long poisoned passages under one fixed context-window policy. How much of the reported gain remains after this control?
  5. The evaluated defenses operate on individual passages, although the attack targets execution control. Please test a controller that blocks semantically repeated queries. Does CyclicRA retain an advantage when this check is configured to preserve clean-task accuracy?
Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to W1 and W2

[W1] Does cycle closure provide additional gains over an open-chain control, and how often is the intended cycle completed?

Cycle closure provides additional resource amplification over the open-chain control, and the intended cycle is completed in roughly half of attacked executions on average across the evaluated settings. We support these findings with an open-chain control that changes only the final dependency and trajectory-level cycle-completion statistics, as detailed below.

[W1] Could the authors change only the final redirection to create an open-chain control?

We replace the original ablation with an open-chain control by redirecting only the final dependency to a previously unused query, while keeping all other query dependencies unchanged. The results are shown in following table.

Table.R1: Coordinated Acyclic (CA) versus CyclicRA under matched document-generation budgets (SQuAD).

ModelMethodTCRΔSTEPSΔCAP HITS (pp)
Qwen3.5-4BCyclicRA31.669.3811.50
CA33.859.745.20
Qwen3.5-9BCyclicRA82.1612.2025.80
CA61.629.009.10
Phi4-14BCyclicRA97.8315.7530.10
CA83.388.9010.60
Gemma3-12BCyclicRA148.5519.0157.70
CA113.5015.9427.00

Table.R1 suggests that cycle closure provides additional resource amplification beyond cross-document coordination, as it enables repeated traversal of the retrieval path rather than a single-pass sequence.

[W1] Please report the fraction of runs that actually complete the intended cycle.

We analyze attacked retrieval trajectories from the main experiments using three metrics: Cycle completion is the percentage of tasks that complete the cycle at least once; Multiple cycle completions measures tasks that complete it at least twice; and Cycle-query revisit rate measures the proportion of tasks that revisit a query in the intended cycle after retrieving at least one different query, capturing recurrence without requiring completion of the full cycle.

Table.R2: Observed cycle completion and cycle-query revisits across victim models and datasets.

ModelDatasetCycle completion (%)Multiple cycle completions (%)Cycle-query revisit rate (%)
Qwen3.5-4BSQuAD50.823.672.3
HotpotQA34.121.568.0
Qwen3.5-9BSQuAD65.940.171.5
HotpotQA46.624.078.0
Phi4-14BSQuAD67.345.585.8
HotpotQA49.433.681.6
Gemma3-12BSQuAD70.450.989.1
HotpotQA55.142.186.7
GPT-5.6-LunaSQuAD56.141.877.6
HotpotQA37.520.459.2

The results show that, on average, half of the tasks complete the intended cycle, with nearly half of these completing multiple cycles. Most remaining tasks still revisit earlier queries along the intended path, demonstrating CyclicRA's ability to induce recurrent retrieval and amplify resource consumption.

[W2] Why should NLI scores between questions predict the agent’s next search?

Our method uses NLI scores as a heuristic to identify subquestions with plausible inferential connections when constructing a retrieval cycle, rather than as a predictor of the agent’s next search. Prior work supports using NLI to assess inferential relations between natural-language texts [1]. Whether these connections translate into the agent’s actual next search also depends on the poisoned documents generated to establish these dependencies and the agent’s evolving execution context. We have clarified this role in Section 4.1 of the revised PDF (pp. 4–5, L215–240).

AuthorsOpenReview ↗

Response to W3, W4 and W5

[W3] Does maximizing the weakest edge outperform the sum-based objective under the same search budget?

The weakest-edge objective does not consistently outperform the sum-based objective, which achieves higher ΔCAP HITS across all four evaluated models. We compare both objectives on 200 randomly selected SQuAD tasks under the same search budget, as detailed below. For this comparison, we replace the sum-based objective in simulated annealing with an objective that maximizes the minimum edge weight:

Smin⁡(Γ)=min⁡{w(q0,c1),w(c1,c2),…,w(ck,c1)}.S_{\min}(\Gamma)=\min\{w(q_0,c_1),w(c_1,c_2),\ldots,w(c_k,c_1)\}. Table.R3: Sum-based versus max–min cycle objectives on 200 SQuAD tasks.

ModelObjectiveTCRΔSTEPSΔCAP HITS (pp)
Qwen3.5-4BSum35.4410.2010
Max–min32.088.139
Qwen3.5-9BSum97.429.1929
Max–min89.129.2524
Phi4-14BSum90.6611.7524
Max–min76.4212.3020
Gemma3-12BSum112.1213.6043
Max–min179.6015.4839

The weakest-edge objective does not consistently outperform the sum-based objective, supporting our original design choice for effective cycle construction.

[W4] Does CyclicRA’s advantage persist with equal-length poisoned passages and a fixed context-window policy?

CyclicRA retains the highest TCR, ΔSTEPS, and ΔCAP HITS on both evaluated models after controlling for passage length and context-window policy. We compare all three methods on 100 SQuAD tasks using the same attack budget, and an 8,192-token sliding history window, as detailed below.

Table.R4: Equal-length poisoned passages with an 8,192-token sliding history window.

ModelMethodTCRΔSTEPSΔCAP HITS (pp)
Qwen3.5-4BOverThink1.30-0.100
STA24.607.408
CyclicRA29.809.809
Qwen3.5-9BOverThink3.600.100
STA46.505.9012
CyclicRA61.2012.5027

Although TCR decreases for all methods under the fixed context-window policy, CyclicRA still achieves the highest TCR on both evaluated models.

[W5] Does CyclicRA retain an advantage against repeated-query detection calibrated to preserve clean-task accuracy?

CyclicRA remains comparatively effective under execution-level repeated-query detection, suggesting that its nonconsecutive query revisits help sustain the attack under this defense. We evaluate LoopGuard on 200 SQuAD attack trajectories across four models, using warning and stopping thresholds of 0.8 and 0.9 calibrated on a separate clean validation set.

Table.R5: LoopGuard warning and interception rates on 200 SQuAD attack trajectories (%).

ModelMethodWarning rate (%)Interception rate (%)
Qwen3.5-4BOverThink7261
STA4210
CyclicRA3713
Qwen3.5-9BOverThink7854
STA5239
CyclicRA4721
Phi4-14BOverThink5942
STA4717
CyclicRA4122
Gemma3-12BOverThink5639
STA2611
CyclicRA226

The results suggest that CyclicRA’s nonconsecutive query revisits are difficult for LoopGuard to detect under the current configuration.

[1] Weir, N., et al. (2024). Enhancing Systematic Decompositional Natural Language Inference Using Informal Logic. Proceedings of EMNLP 2024, 9458–9482.

Sorry for the late reply. I have read the rebuttal and appreciate the additional experiments, but I still have two remaining concerns.

  1. Describing NLI as a heuristic does not fully explain why question-pair entailment scores help construct effective retrieval cycles.
  2. Could you explain why related attacks such as CREEP [1] and LoopTrap [2] were not included as baselines?

I welcome further clarification and corrections to any misunderstandings. For now, I lean toward maintaining my original score.


[1] Inference Cost Attacks for Retrieval-Augmented Large Language Models.

[2] LoopTrap: Termination Poisoning Attacks on LLM Agents.

Reviewer_oSh7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies denial-of-efficiency attacks on agentic retrieval-augmented generation. An attacker knows a target query and inserts documents into the retrieval corpus, without modifying the victim model, retriever, or controller. CyclicRA generates intermediate questions, scores directed question pairs using an NLI model, searches for a high-scoring cycle with simulated annealing, and evolves a document for each transition. These documents present apparent information dependencies intended to keep the victim retrieving.

Critical strengths
  1. The resource-consumption objective is consequential and clearly separated from answer manipulation. The attack operates through corpus insertion under an explicitly query-aware threat model (§3, pp. 3–4). Studying how retrieved material prolongs execution is a useful extension of corpus-poisoning evaluation, particularly because the paper measures retrieval steps and budget exhaustion alongside tokens.

  2. The reported execution overhead is consistent across the evaluated models. CyclicRA has the largest increase in retrieval steps in all ten model–dataset settings and the highest TCR in seven (Table 1, p. 7). Figure 4 also shows higher TCR and step increases than the no-loop variant in all eight tested settings. This consistency supports investigating coordinated document poisoning, even though it does not by itself establish the precise mechanism.

Critical weaknesses

Please refer to questions.

Questions for authors
  1. What fraction of attacked executions completes the intended cycle and traverses it more than once? How are semantically equivalent queries identified, and how much additional cost comes from recurring cycles versus nonrecurring detours? Can representative anonymized trajectories substantiate this?

  2. Does the advantage persist against a coordinated acyclic control with matched document count, length, and generation/optimization budget? Is NLI score predictive of observed transitions, and is the evolutionary optimization budget matched for the no-loop variant?

3 comments
AuthorsOpenReview ↗

Response to Q1

[Q1] To what extent does the intended retrieval cycle recur during attacked executions and contribute to the additional resource consumption?

The intended retrieval cycle occurs and is repeatedly traversed, with cycle-related retrievals accounting for 60.81% of retrieval steps whose queries are absent from the corresponding clean executions. We support this conclusion with semantic query matching, trajectory-level statistics, and comparisons with clean executions, as detailed below.

[Q1] How are semantically equivalent queries identified?

We encode queries as sentence embeddings using all-MiniLM-L6-v2 and compute their cosine similarity. We set the similarity threshold to 0.9 and use GPT-4o-mini to assess borderline cases.

[Q1] What fraction of attacked executions completes the intended cycle and traverses it more than once?

We analyze attacked retrieval trajectories from the main experiments using three metrics: Cycle completion is the percentage of tasks that complete the cycle at least once; Multiple cycle completions measures tasks that complete it at least twice; and Cycle-query revisit rate measures the proportion of tasks that revisit a query in the intended cycle after retrieving at least one different query, capturing recurrence without requiring completion of the full cycle.

Table.R1: Observed cycle completion and cycle-query revisits across victim models and datasets.

ModelDatasetCycle completion (%)Multiple cycle completions (%)Cycle-query revisit rate (%)
Qwen3.5-4BSQuAD50.823.672.3
HotpotQA34.121.568.0
Qwen3.5-9BSQuAD65.940.171.5
HotpotQA46.624.078.0
Phi4-14BSQuAD67.345.585.8
HotpotQA49.433.681.6
Gemma3-12BSQuAD70.450.989.1
HotpotQA55.142.186.7
GPT-5.6-LunaSQuAD56.141.877.6
HotpotQA37.520.459.2

The results show that, on average, half of the tasks complete the intended cycle, with nearly half of these completing multiple cycles. Most remaining tasks still revisit earlier queries along the intended path, demonstrating CyclicRA's ability to induce recurrent retrieval and amplify resource consumption.

[Q1] How much additional cost comes from recurring cycles versus nonrecurring detours?

Cycle-related retrievals account for 60.81% of retrieval steps whose queries are absent from the corresponding clean executions, while other detours account for 39.19%. This indicates that the intended cycle is the main source of attack-induced query deviations, supporting CyclicRA’s effectiveness in steering retrieval along the constructed path and increasing retrieval overhead. We obtain this breakdown by comparing attacked trajectories with their clean counterparts and classifying these additional steps into cycle-related retrievals and other detours.

[Q1] Can representative anonymized trajectories substantiate this?

We have included a representative retrieval trajectory in Appendix E.2 of the revised PDF (pp. 22–24, L1158–1265).

AuthorsOpenReview ↗

Response to Q2

[Q2] Does cycle closure provide additional gains under matched budgets, and what role does NLI play in constructing the cycle?

CyclicRA retains its advantage over the Coordinated Acyclic control in most evaluated settings under matched document counts, passage-length constraints, and generation/optimization budgets. NLI scores serve as a heuristic for selecting plausible connections between subquestions; actual query transitions also depend on the poisoned documents that establish these dependencies. Below is our detailed response.

[Q2] Does the advantage persist against a coordinated acyclic control with matched document count, length, and generation/optimization budget?

We replace the original no-loop ablation with a Coordinated Acyclic control (CA), redirecting only the final dependency to a previously unused query instead of closing the cycle.

Table.R2: Coordinated Acyclic (CA) versus CyclicRA under matched document-generation budgets (SQuAD).

ModelMethodTCRΔSTEPSΔCAP HITS (pp)
Qwen3.5-4BCyclicRA31.669.3811.50
CA33.859.745.20
Qwen3.5-9BCyclicRA82.1612.2025.80
CA61.629.009.10
Phi4-14BCyclicRA97.8315.7530.10
CA83.388.9010.60
Gemma3-12BCyclicRA148.5519.0157.70
CA113.5015.9427.00

Table.R2 suggests that cycle closure provides additional resource amplification beyond cross-document coordination, as it enables repeated traversal of the retrieval path rather than a single-pass sequence.

[Q2] Is NLI score predictive of observed transitions?

Our method uses NLI scores as a heuristic to identify subquestions with plausible inferential connections when constructing a retrieval cycle, rather than as a predictor of observed query transitions. Prior work supports using NLI to assess inferential relations between natural-language texts [1]. Whether these connections translate into actual query transitions also depends on the poisoned documents generated to establish these dependencies and the agent’s evolving execution context. We have clarified this role in Section 4.1 of the revised PDF (pp. 4–5, L215–240).

[Q2] Is the evolutionary optimization budget matched for the no-loop variant?

Both the original CyclicRA (w/o loop) variant and the Coordinated Acyclic control match CyclicRA in poisoned-document count, passage-length constraints, and evolutionary optimization budget. Further details on the Coordinated Acyclic control are provided in Section 5.3 of the revised PDF (p. 9, L468–477).

[1] Weir, N., et al. (2024). Enhancing Systematic Decompositional Natural Language Inference Using Informal Logic. Proceedings of EMNLP 2024, 9458–9482.

Reviewer_oSh7OpenReview ↗

Thanks for the response

Thanks for the authors' response. My questions addresses. And I would like to keep my score.

Comments on the paper

1 comment
Area_Chair_UVeXOpenReview ↗

Reminder: Discussion Period Closes in 24 Hours

Dear Reviewers,

This is a gentle reminder that the discussion period will close in 24 hours.

If you have not yet done so, please review the authors’ responses to your comments and participate in the discussion. Please let the authors know whether their responses have addressed your concerns or whether you have any remaining questions.

Thank you again for your time and contribution to the review process.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.