acceptodds
Under review as a conference paper at ICLR 2027

AttackRecBench: Benchmarking Item Promotion Attacks on Agentic Recommender Systems

Abstract

Large language models are transforming recommender systems from one-shot rankers into agents that retrieve external information and maintain state across multi-step decisions. Existing recommendation-attack benchmarks do not systematically evaluate the new attack surfaces introduced by these agentic capabilities. In particular, attacker-controlled item titles, descriptions, and public reviews can manipulate evidence retrieval, reasoning, memory, and final rankings. We introduce AttackRecBench, a benchmark for item-promotion attacks on agentic recommender systems. It comprises 1,500 candidate-set ranking tasks from Amazon, Goodreads, and Yelp across five recommendation scenarios. The benchmark includes eleven attacks spanning content rewriting, indirect prompt injection, decision-bias manipulation, and working-memory poisoning, evaluated across five recommender-agent architectures. Attackers can modify only seller- or user-controlled content, without access to model parameters, user histories, system instructions, or candidate sets. Paired clean and adversarial evaluation measures target-item promotion and recommendation utility. In our primary evaluation setting, prompt injection is the strongest promotion mechanism, increasing the top-three exposure rate by 17.7–40.6 percentage points across the three domains. Attack effectiveness is nevertheless conditional: the strongest individual attack changes across recommendation scenarios and agent architectures, while responses to the memory-oriented attack condition vary markedly across architectures. We further evaluate three lightweight inference-time defenses. All three reduce the top-three exposure rate of the strongest contextual prompt-injection attack by 25.6–29.7 percentage points, but their security and utility effects vary across attack families; no defense dominates overall.

Primary area:
datasets and benchmarks
Keywords:
agentic recommender systems, recommendation security, item-promotion attacks

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_5eLAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates item-side security risks in agentic recommender systems, focusing on whether attackers can manipulate recommendation rankings by modifying publicly available item information, such as titles, descriptions, and reviews. To examine this problem, the authors introduce AttackRecBench, a benchmark designed to systematically compare different attack strategies under consistent evaluation settings. The framework assesses attack effectiveness from two perspectives: the promotion of targeted items and the preservation of overall recommendation quality. The experiments span three data sources, Amazon, Goodreads, and Yelp, and consider multiple attack techniques, recommendation workflows, and underlying language models. In addition, the paper evaluates several lightweight defense mechanisms. The findings suggest that prompt injection is particularly effective under the main experimental setting, although attack performance and defense effectiveness vary depending on the specific configuration.

Critical strengths
  1. The paper investigates target-item promotion through modifications to public content, without requiring access to user histories, victim model internals, or victim feedback. The experimental results demonstrate that attackers can significantly influence recommendation outcomes even with limited control. This reveals potential security vulnerabilities in real-world recommender systems and highlights the practical significance of the study.
  2. The paper jointly evaluates defense effects on target-item promotion and recommendation quality, showing that security benefits and utility costs vary across attacks. This provides empirical evidence about security–utility trade-offs and makes the benchmark useful for comparing and selecting defense strategies.
Critical weaknesses
  1. The paper states: “An agent that ignores useful evidence may appear robust even though it is simply a weak recommender.” Yet Figure 2(b) reports relative changes without complete absolute clean/attacked recommendation quality. AdaptiveTool’s 11.6% source exposure and small attack effects could reflect either effective selective retrieval or insufficient evidence use. MACRec also uses a fixed six-stage workflow without Searcher, Interpreter, or dynamic retrieval, while evidence access and call budgets differ across architectures. Capability validation and mechanism controls are therefore needed to interpret these security–utility differences.
  2. The paper defines memory poisoning as attacked content entering writable working memory, yet Table 3 provides no WMP-specific memory-carry rate, leaving unclear how many WMP executions satisfy this definition. Both memory architectures also retain the original attacked evidence during ranking, so target promotion alone cannot establish the effect of memory contamination. Moreover, the evaluated memory is task-local and excludes existing cross-user collaborative and retrieval-based designs.
Questions for authors
  1. Could the authors provide absolute clean/attacked NDCG@3 and ER@3 by architecture and dataset for Figure 2(b), alongside actual model and tool calls, to clarify whether AdaptiveTool’s small attack effects coexist with competitive recommendation quality?
  2. Which attacks and denominators underlie Table 3’s Memory metric, and how is memory carry identified? For MemoryAgent and Agent4Rec, please report WMP execution and memory-carry counts, clarifying whether the retained content includes the attacker-induced favorable association rather than ordinary item information.
  3. Could the authors report AdaptiveTool’s paired clean/attacked access rates to the same target source, and distinguish unexposed executions that submit rankings before exhausting the tool budget from those that exhaust it? This would clarify how retrieval decisions and budget constraints relate to the low exposure rate.
4 comments
AuthorsOpenReview ↗

PART I: Response to Reviewer 5eLA — Absolute Metrics, Call Budgets, and Architectural Controls (Q1)

Thank you for recognizing our black-box threat model (which requires no access to user histories, model parameters, or feedback) and our empirical analysis of lightweight defense trade-offs. We organize our response across three parts (Part I: Q1; Part II: Q2; Part III: Q3) based on our full execution logs (Appendix B.2) and new controlled experiments.


Q1: Absolute performance metrics corresponding to Figure 2(b), actual call budgets, and architectural capability controls.

A1: Thank you for raising these important points, which highlight two key issues: testing the "weak-recommender" explanation via absolute metrics and actual call counts, and clarifying capability controls across architectures.

1. Absolute Performance and Actual Call Overhead (Appendix D.2, Table 4) In Appendix D.2 of the revision, we added Table 4 reporting absolute Clean/Attacked NDCG@3, absolute Clean/Attacked ER@3, and executed model calls (MC) and data-tool calls (TC) across all five architectures on Amazon, Goodreads, and Yelp (under configured limits of 10 tool calls and 30 model turns).

  • Ruling out the weak-recommender effect: The Introduction's remark that an agent ignoring useful evidence can appear robust simply due to weak recommendation quality was meant as a methodological principle (motivating paired clean/attacked runs and concurrent NDCG@3 reporting), not a claim about a specific architecture. Table 4 directly tests this hypothesis. Under clean conditions, AdaptiveTool achieves 37.4% NDCG@3, matching or slightly exceeding fellow multi-step baselines MACRec-style (36.1%) and Agent4Rec (36.8%), while trailing BaseAgent (45.2%), which ingests the entire candidate pool in one pass. We do not claim stronger recommendation utility than full-context pre-fetching. Under attack, AdaptiveTool maintains 36.8% NDCG@3 and 12.2% ER@3, virtually unchanged (+0.2 percentage points) from its 12.0% clean baseline, whereas BaseAgent's ER@3 jumps from 11.7% to 20.9% (+9.2 points). As detailed in Q2 (Part II), AdaptiveTool's small attack effect stems from rarely reading item descriptions (3.7% clean vs. 3.5% attacked on the 300-task subset), not from resisting payloads once read: when descriptions are forcibly exposed, DRI's ER@3 climbs to 78.3% (vs. 13.7% standard). Our claim is therefore that selective evidence access reduces exposure, not that AdaptiveTool is inherently robust or filters payloads.
  • Actual call overhead: Recomputed from raw logs over valid rankings (length 20, no duplicates), BaseAgent, MemoryAgent, and MACRec-style execute fixed model calls matching their stage counts (1.00, 2.00, and 6.00, with 0.00 tool calls). AdaptiveTool averages 3.30–4.83 model calls (3.95 overall) and 2.30–3.83 data-tool calls (2.95 overall), well below the 10-tool / 30-turn caps, with clean and attacked averages differing by at most 0.01. By contrast, Agent4Rec averages only 1.35 model calls and 0.15 data-tool calls (0.02–0.28 across datasets), rarely invoking retrieval, which explains why its attacked ER@3 (19.0% macro-average) stays close to non-retrieval baselines.

2. Architectural Capability and Mechanism Control Boundaries We agree that these implementation details belong in the main paper and have moved them from Appendix B.1 into Sections 4.1.2 and 4.2.2. Specifically, MACRec-style adapts MACRec to our closed-set candidate reranking setting via a fixed six-stage workflow (manager plan → user analyst → item analyst → manager draft → reflector → manager final). We intentionally exclude the upstream framework's Searcher, Interpreter, and dynamic retrieval loops for two reasons: (i) in closed-set reranking where candidate metadata is already structured and bounded, external search/code-execution tools are unnecessary, and (ii) keeping MACRec-style on full-context prefetching isolates the effect of multi-role deliberation and reflection without confounding it with on-demand tool retrieval (tested separately in AdaptiveTool and Agent4Rec). Accordingly, we explicitly label this variant MACRec-style, avoid extrapolating to the full upstream framework with dynamic Searcher enabled, and qualify cross-architecture differences as reflecting the joint effect of evidence access channels and state/collaboration budgets rather than a single isolated factor.

AuthorsOpenReview ↗

PART II: Response to Reviewer 5eLA — Paired Target Source Access Rate and Budget Exhaustion Analysis (Q2)

Q2: Paired target source access rate and tool budget exhaustion analysis for AdaptiveTool.

A2: Following your suggestion, we re-ran AdaptiveTool under paired clean and attacked conditions on the 300-task subset across all 11 attacks (3,299 valid attacked runs) and analyzed the retrieval logs.

1. Paired Target Source Access Rate For the 9 description-based attacks, AdaptiveTool accessed the target item's description in 3.7% (11/300) of clean runs and 3.5% (95/2,699) of attacked runs. Across all 2,699 (task × attack) pairs, 68 accessed the target description only when clean, 64 only when attacked, 31 in both, and 2,536 in neither. Injected text does not alter AdaptiveTool's propensity to inspect target descriptions.

For the remaining two attacks: LP modifies item titles, which are always visible in the initial candidate list (300/300 attacked runs exposed). PRI modifies reviews; the injected review appeared in tool outputs in 8 of 300 runs (2.7%, with the target reaching top-3 in 5 of those 8) and did not appear in the remaining 292 runs (where the target reached top-3 in 41; given the small sample, we report these counts descriptively).

2. Termination Reasons and Budget Breakdown for Unexposed Runs (87.8%) Across the 3,299 attacked runs, the target source was exposed in 403 runs (12.2%, close to the 11.6% full-dataset rate in Table 3: 300 LP, 95 description-based, and 8 PRI). Among the 2,896 runs (87.8%) where the target remained unexposed:

  • 98.3% (2,848 runs) voluntarily called submit_ranking before exhausting the 10-call tool budget, using ~2.7 data-tool calls on average;
  • Only 1.7% (48 runs) exhausted all 10 tool calls before final submission, and zero runs hit the 30-turn model interaction limit.

Thus, unexposed runs happen mostly because the agent stops retrieving early, not because it runs out of budget. Importantly, this does not mean AdaptiveTool is inherently robust to attacks. Across the 9 description-based attacks, ER@3 is 40.0% (38/95) when the target description is read versus 13.3% (346/2,604) when unread (an observational comparison). When we force candidate descriptions into context, prompt-injection attacks spike in ER@3: DRI rises from 13.7% to 78.3%, COI from 13.3% to 76.3%, and BSI from 14.7% to 40.0%, whereas the decision-bias attack BPI (authority/brand-prestige cues) barely moves (13.7% vs. 13.0% standard). The low overall attack effect therefore comes from reduced exposure via selective evidence access, not robustness against prompt injection.

AuthorsOpenReview ↗

PART III: Response to Reviewer 5eLA — Memory Metric Definitions, WMP Carryover, and Counterfactual Check (Q3)

Q3: Table 3 memory metric definitions, WMP-specific carryover rate, and counterfactual memory-only check.

A3: Thank you for highlighting these three issues regarding Table 3's memory metrics. We fully agree that WMP-specific carryover rates were not isolated, retaining original attacked evidence during final ranking prevented separating memory pollution from direct evidence exposure, and our evaluation covers task-local working memory. We address each below.

1. Table 3 Denominators and Carryover Criteria The Memory column in Table 3 (MemoryAgent 0.7%, Agent4Rec 2.6%) uses total attacked runs across all 11 attacks per architecture as its denominator (Appendix D.1) and applies only to MemoryAgent and Agent4Rec, which maintain explicit writable memory slots (marked – for the other three architectures). Following Appendix D.1, "memory carryover" is logged whenever content from an attacked field is written into a working-memory slot. Because this structural criterion does not distinguish attacker-induced endorsements from benign item attributes, we added a content-level inspection below.

2. WMP-Specific Execution and Carryover Statistics (Appendix D.3, Table 5) Re-running WMP alone on the 300-task subset (Appendix D.3, Table 5), memory carryover occurs in 11/300 runs (3.7%) for MemoryAgent and 10/299 runs (3.3%) for Agent4Rec. Because these rates use WMP-only runs as the denominator whereas Table 3 pools all 11 attacks, the two sets of numbers are not directly comparable.

To inspect what was actually written across all 21 carryover cases, we checked both verbatim copying (any contiguous 5-word sequence from the attack text appearing verbatim in memory: 5/11 for MemoryAgent, 1/10 for Agent4Rec) and semantic adoption via unified manual annotation (whether memory explicitly names the target item and adopts attacker-induced claims such as "strong match," "reliable," or "prioritize"). Under manual review, adversarial content carryover holds in 10/11 cases for MemoryAgent but only 1/10 for Agent4Rec (the other 9 logged entries for Agent4Rec never mention the target item and contain only generic user-preference summaries).

3. Counterfactual Memory-Only Check As noted in Appendix D.1, descriptive carryover rates do not by themselves prove that memory pollution drives ranking shifts. To isolate the memory pathway, we ran a counterfactual check where, after memory is written, final ranking retains only the poisoned memory while restoring all item evidence to clean text.

  • MemoryAgent (300-task subset): With poisoned memory alone, target ER@3 under WMP reaches 20.1% (vs. 12.8% clean baseline and 33.3% standard attack), showing that the memory pathway alone produces a measurable exposure gain, though smaller than the full attack.
  • Agent4Rec (60 item cold-start tasks): Because the remaining 240 tasks rarely trigger the two-stage "write memory → re-rank" flow, the counterfactual is only meaningful on these 60 tasks. Here, WMP ER@3 is 11.7% (clean), 16.9% (standard attack), 8.5% (clean evidence + poisoned memory), and 11.7% (fresh-session control). Paired differences are statistically indistinguishable from noise (memory-only vs. control: −3.4, 95% CI [−10.2, +1.7]; standard attack vs. clean: +5.1, 95% CI [−1.7, +13.6]; n ≈ 59–60). We find no discernible causal effect through the memory pathway in Agent4Rec and make no claim on whether memory is its primary attack channel.
  • Scope boundary: Our study evaluates task-local working memory; cross-user collaborative memory and retrieval-based long-term memory are outside our scope, and we explicitly refrain from extrapolating our findings to those settings.

Dear Reviewer 5eLA,

Thank you very much again for your careful reading and the precise, diagnostic questions — they have substantially improved the paper.

As the discussion period draws to a close today, we wanted to check whether any of your questions remain unaddressed, or whether anything in our responses or the revised manuscript has raised new confusion. If so, please let us know — we will respond and resolve it right away, while there is still time.

Thank you again for your time and consideration.

Reviewer_g7C1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Existing recommendation-attack benchmarks do not systematically evaluate the new attack surfaces introduced by the agentic capabilities. The paper introduces a controlled benchmark of 11 item-promotion attacks spanning four manipulation mechanisms and three attacker-writable sources over 1,500 AgentRecBench tasks from three datasets and five recommendation scenarios. The paper designs a paired evaluation protocol that fixes the promotion target and underlying evidence across conditions, compares five agent architectures, and jointly measures top three exposure, rank improvement, and recommendation utility. The paper provides dataset-, scenario-, and architecture-level analyses of attack effectiveness and evaluate three representative inference-time defenses, revealing context- and mechanism-dependent security–utility trade-offs.

Critical strengths

(1) The paper is related to an emerging topic. The developed benchmark is very helpful for the evaluation of a real-world challenge. This is very helpful for the researchers working on this area.

(2) The evaluation and analysis is very comprehensive, including various datasets, scenarios, agent architectures and metrics.

Critical weaknesses

(1) It might be helpful to add evaluations using model with varying capabilities, to understand whether the observations on the benchmark datasets are sensitive to the model capabilities. For example, it might be helpful to evaluate more advanced models (e.g., GPT-6.1 Sol) or open-source models with smaller model sizes (e.g., 1.7B or 4B).

(2) It might be helpful to discuss and highlight the distinguishing features of the developed benchmark, compared to existing benchmarks, so the positioning of this paper is more clear.

Questions for authors

(1) Are the observations on the benchmark datasets sensitive to the model capabilities? For example, it might be helpful to evaluate more advanced models (e.g., GPT-6.1 Sol) or open-source models with smaller model sizes (e.g., 1.7B or 4B).

(2) Might it be feasible to discuss and highlight the distinguishing features of the developed benchmark (e.g., by a table where the columns are the features), compared to existing benchmarks?

2 comments
AuthorsOpenReview ↗

Response to Reviewer g7C1 — Model Capability Sensitivity (W1/Q1) and Benchmark Comparison (W2/Q2)

Thank you for your positive assessment of the timeliness and practical security significance of our work, as well as the comprehensiveness of our datasets, recommendation scenarios, agent architectures, and evaluation metrics. We have incorporated both of your suggestions into the revised manuscript as detailed below.

W1 & Q1: Sensitivity to model capability.

A1: We agree that examining sensitivity across model scales is important. In addition to GPT-OSS-120B used in our main experiments, Appendix D.6 of the revised manuscript reports results on Qwen-72B-Instruct. Regarding smaller open models and frontier proprietary models, our findings and scope boundaries are as follows:

  • Smaller parameter models: We evaluated Qwen3-4B-Instruct-2507 and Qwen2.5-7B-Instruct on the 300-task subset (BaseAgent, temperature 0). Under our strict output protocol (exact 20-item rankings with no duplicates), valid ranking rates without fine-tuning were only 28.6% and 7.8%, respectively (most outputs contained duplicate items), falling well below our pre-defined usability threshold (≥90% valid rate on clean runs and ≥85% on attacked runs). Without task-specific fine-tuning, these smaller models cannot reliably support our agentic recommendation benchmark; accordingly, our evaluation focuses on larger models capable of stable structured ranking, leaving fine-tuning of small models and their security behavior to future work.
  • Frontier proprietary models: Our benchmark is scoped to mainstream open-weight models (GPT-OSS-120B, Qwen-72B-Instruct, etc.). Conducting full-matrix evaluations across all architectures, attacks, and tasks on the latest frontier proprietary models (e.g., GPT 6.1 Sol) incurs prohibitive API costs and falls outside the scope of this study, which we explicitly note as a direction for future work.

W2 & Q2: Multi-dimensional comparison with existing benchmarks and attack studies (Appendix D.5, Table 8).

A2: Following your suggestion, we added a multi-dimensional comparison against existing benchmarks and attack studies in Appendix D.5 (Table 8) and referenced it in Related Work. Entries are compiled from each paper's main text, where — denotes that the aspect is not specified in the main text rather than necessarily absent:

WorkSettingAttacker InputEvaluated ArchitecturesDefensesPropagation Tracking
AgentRecBenchAgentic recommendation accuracy (HR@N), no attacks—8 single-agent systems (multi-agent as future work)——
Stealthy AttackLLM recommender modelsTest-time item text modification (black-box)4 LLM recommendersYes (paraphrasing)—
Rank Anything FirstLLM rerankers over item categoriesOptimized adversarial token sequences4 LLM rerankers——
Tang et al. (Bias)Single-step candidate selection10 bias types injected into candidate textMultiple LLMs (single-step selection)Yes—
DrunkAgentMemory-based agentic recommendationBlack-box modification of target description3 memory-based agentsYes—
Visual InceptionAgentic recommendation with long-term memory (simulated e-commerce)Poisoned user-uploaded images (black-box)1 agent environment, multiple LLM backendsYes (proposed & baselines)—
AttackRecBench (Ours)5 agent architectures, unified paired protocol4 categories, 11 attacks (one public field each)53 defensesYes (retrieval / memory / relay)

Thank you very much for the detailed responses and comprehensive comparisons. I think my questions have been addressed. I would keep the current positive evaluation.

Reviewer_pafcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces AttackRecBench, a benchmark for evaluating item-promotion attacks on agentic recommender systems. The benchmark comprises 1,500 recommendation tasks from Amazon, Goodreads, and Yelp, covering five recommendation scenarios, eleven attack methods, and five agent architectures. Its threat model is reasonably realistic: attackers have no access to model parameters, system instructions, user histories, or candidate sets and can only modify legitimately controllable content, such as item titles, descriptions, and public reviews. Through paired comparisons between clean and attacked settings, the experiments jointly evaluate the target item’s top-three exposure rate, rank improvement, and recommendation utility.

Critical strengths
  1. The topic is practically relevant, and the paper systematically reveals the new attack surfaces introduced by retrieval, memory, and multi-step reasoning in agentic recommender systems.

  2. The experimental design is comprehensive and carefully controlled. By fixing user histories, candidate sets, target items, and underlying evidence, the authors conduct paired comparisons between clean and attacked settings. They jointly report attack effectiveness, rank changes, and recommendation utility, thereby avoiding the potential misinterpretation of weak recommendation capability as strong robustness.

Critical weaknesses
  1. The comparability of the different attack methods considered in the paper is difficult to assess, as attack success may be driven by differences in prompt wording, historical memory, or other implementation details. The current experiments only demonstrate that the selected explicit ranking instructions manipulate the model more effectively than the selected simple marketing statements. However, they do not provide sufficient evidence to support the broader conclusion that prompt injection is generally a stronger mechanism than other types of attacks.

  2. The experiments essentially evaluate offline reranking over a fixed candidate set and therefore do not cover the major agentic attack surfaces emphasized in the paper’s motivation. Experiments on online or deployed recommendation platforms should be included to validate whether the reported vulnerabilities remain effective in more realistic settings.

Questions for authors

See the weaknesses

3 comments
AuthorsOpenReview ↗

PART I: Response to Reviewer pafc — Attack Comparability, Controls, and Sensitivity Analyses (W1)

Thank you for recognizing the practical value of uncovering retrieval, memory, and multi-step reasoning attack surfaces in agentic recommenders, as well as our paired evaluation design (fixed user history, candidate set, and evidence) and joint reporting of exposure, rank shift, and recommendation utility. We address your two core concerns across two parts (Part I: W1 controls and sensitivity analyses; Part II: W1 claim calibration and W2) through existing controlled baselines, new sensitivity experiments, and calibrated claim boundaries.


W1 (Part 1/2): Attack comparability and the boundary of the claim that prompt injection generally outperforms other mechanisms.

A1: Thank you for this rigorous critique. Below we detail (1) existing strict variable controls in our benchmark and (2) new paraphrase and length-matched sensitivity experiments, followed by our revised manuscript claim boundaries in Part II.

1. Existing Variable Controls in the Benchmark As detailed in Appendices B.2 and C, every clean, attacked, and defended run executes in a completely reset fresh session: tool caches, writable working memory, and inter-agent messages are wiped between conditions, while user histories, candidate orderings, and evidence for the other 19 candidates are strictly frozen. Cross-attack differences are therefore free from session-state or memory cross-contamination. Furthermore, among the 11 attacks, indirect prompt injections (DRI, COI, BSI) and decision-bias attacks (BPI, SPI, SUI) modify the exact same writable field (item Description) at the exact same position (appending a 30–171 character payload while keeping the original description intact as prefix; Appendices A.2–A.3). Thus, the large gap between COI/BSI (26.2%–54.6% ER@3 in the main setting) and BPI/SPI/SUI (10.3%–14.8%, near clean baselines) is not confounded by target field, insertion position, or session state.

2. Paraphrase Sensitivity and Length-Matched Controls (Appendix D.4, Tables 6–7) To rule out wording artifacts (strong instruction templates vs. weak marketing phrases), we ran controlled sensitivity experiments on the 300-task cross-dataset subset, reporting all variant texts and ER@3 scores in Appendix D.4 (Tables 6 and 7):

  • Paraphrase sensitivity on BaseAgent (Table 6): We constructed 3 semantically equivalent paraphrases per attack type for prompt injection (DRI, COI, BSI; 9 variants) and decision bias (BPI, SPI, SUI; 9 variants). Prompt-injection variants average 57.7% ER@3 (range 14.9%–94.5%) compared to 13.3% (range 11.1%–15.2%) for decision-bias variants. While the family-level advantage is consistent, prompt injection is highly sensitive to phrasing (e.g., one BSI variant drops to 14.9%, comparable to decision bias). Accordingly, we claim a family-average difference rather than superiority of every individual template.
  • Length-matched control on BaseAgent (Table 6): Lengthening marketing payloads to ~127 and 170 characters while shortening instruction payloads to ~22 and 27 characters yields 32.1% and 35.4% ER@3 for short instructions versus 17.9% and 14.6% for long marketing text. Within this length range, text length alone does not explain the performance gap.
  • Cross-agent replication on MemoryAgent and Agent4Rec (Table 7): On MemoryAgent (first two paraphrase sets P1–P2, GPT-OSS-120B, 300-task subset), prompt-injection variants average 37.0% ER@3 (range 13.2%–91.8%; original 27.9%) versus 13.1% (range 11.5%–14.9%; original 11.7%) for decision bias, again showing strong within-family phrasing variance (COI P1: 13.2% vs. BSI P1: 91.8%). On Agent4Rec (all 3 paraphrase sets for DRI, COI, BSI, and BPI; GPT-OSS-120B, 300-task subset; SPI/SUI omitted), ER@3 results follow the same pattern: DRI original 36.1% vs. variants 54.4% / 35.1% / 36.1% (P1 vs. original: +19.2, 95% CI [+13.7, +25.3]; P2/P3 indistinguishable from original); COI original 24.0% vs. 16.5% / 29.5% / 59.3%; BSI original 36.5% vs. 95.2% / 17.8% / 57.5%; and BPI original 12.3% vs. 13.0% / 15.2% / 13.7% (all 95% CIs of differences include 0). Length controls on Agent4Rec also match BaseAgent: short instructions (22/27 chars) achieve 30.3% and 31.0% ER@3 (DRI original 36.1%), versus 12.1% and 12.3% for long marketing text (BPI original 12.3%).
AuthorsOpenReview ↗

PART II: Response to Reviewer pafc — Claim Boundaries (W1) and Offline vs. Deployed Systems (W2)

W1 (Part 2/2): Calibrating Claim Boundaries in the Manuscript We fully agree that finite representative attack templates should not be generalized into universal laws. We revised the Abstract, Section 1, Section 4.2.1, and Section 5 to explicitly scope this finding to our main evaluation setting (single-field modification on public item fields) and to family averages: Section 4.2.1 now explicitly notes that individual prompt-injection payloads depend strongly on phrasing (Appendix D.4). Under these scoped conditions, prompt-injection attacks (DRI, COI, BSI) achieve higher average ER@3 than content-rewriting and decision-bias attacks, with BSI (Amazon, Goodreads) or COI (Yelp) being strongest per dataset in the main setting, while relative performance varies across scenarios and architectures. In Limitations, we further state that our threat model excludes adaptive, multi-source, and cross-session attacks, and we make no claims on relative attack rankings in those regimes.


W2: Offline fixed-candidate-set reranking versus online/deployed agentic systems.

A2: Thank you for highlighting the gap between offline evaluation and live deployment. We agree that validation on live platforms would further strengthen external validity, as noted in our original Limitations. However, live injection is precluded by ethics and compliance constraints, and our paired benchmark captures multi-step agentic attack surfaces beyond one-shot static reranking:

1. Ethical and Methodological Reasons Against Live Platform Injection Injecting fabricated reviews or prompt-injection payloads into live commercial services (e.g., Amazon, Yelp) violates platform Terms of Service, research ethics standards, and the responsible-research boundaries in our Ethics and Responsible Release section (which prohibits automated attack deployment on production systems). Methodologically, isolating attack-induced rank promotion (ER@3 / Rank Boost) from baseline recommendation utility (NDCG@3) requires executing the exact same task twice (once clean and once attacked) under identical candidate sets and evidence. Such controlled counterfactual pairing is infeasible in live production traffic and is enabled by our AgentRecBench-based design.

2. Agentic Attack Surfaces Captured by the Benchmark Aside from BaseAgent (included as a direct-exposure reference), the other four architectures are multi-step agentic workflows rather than single-pass static rerankers: AdaptiveTool and Agent4Rec autonomously decide which of 7 read-only data tools to query (up to 10 tool calls and 30 model turns); MemoryAgent and Agent4Rec maintain explicit read/write working memory; and MACRec-style executes a six-stage multi-role deliberation and reflection pipeline. Table 3 quantifies how selective retrieval restricts AdaptiveTool's target-source exposure to 11.6%, whereas multi-role message passing propagates injected content across roles in 88.4% of MACRec-style runs. Moreover, the benchmark uses real Amazon, Goodreads, and Yelp item and review texts, modifying only a single seller- or user-editable field. While still an offline evaluation rather than a production deployment, it directly models these intermediate agentic mechanisms.

3. Supporting Empirical Evidence and Explicit Scope Boundaries Our experiments directly substantiate these agentic mechanisms: for dynamic retrieval, AdaptiveTool autonomously selects among 7 read-only tools (averaging 2.95 calls out of 10, with 98.3% of unexposed runs terminating via voluntary early submission; see also our response to Reviewer 5eLA, Q2), and forcibly exposing candidate descriptions raises DRI ER@3 from 13.7% to 78.3%, confirming that dynamic retrieval governs payload exposure. For task-local memory and multi-role propagation, we report WMP memory carryover counts (Appendix D.3, Table 5) and the MemoryAgent counterfactual check (clean evidence + poisoned memory yields 20.1% ER@3 vs. 12.8% clean baseline and 33.3% standard attack).

Finally, we explicitly delineate our evaluation scope: we do not inject attacks into live online platforms, and large-scale candidate retrieval (hundreds of candidates in recall-plus-rerank pipelines) as well as cross-session persistent memory lie outside our current scope. While Limitations already notes that adaptive, multi-source, and cross-session attacks are excluded, the revised manuscript explicitly identifies retrieval scaling and cross-session persistent memory as key directions for future work.

Dear Reviewer pafc, thank you very much for your thorough and critical review, and for recognizing the practical relevance of the topic and the careful control of our paired design. We have reflected carefully on your two central concerns — that our prompt-injection claim was broader than the evidence from the selected attacks could support, and that the offline, fixed-candidate evaluation leaves deployed settings unvalidated. Our rebuttal sought to respond to both, with new paraphrase and length-matched sensitivity analyses (Appendix D.4) and an explanation of the constraints that keep us from experimenting on live production platforms; and we have since revised the manuscript once more (uploaded early this morning) so that the claim is stated in family-average terms throughout, with family-average ER@3 gains of 11.5, 21.6, and 29.9 points now reported alongside the strongest single attack's 17.7, 29.3, and 40.6. As the discussion period draws to a close today, please do not hesitate to let us know if anything remains unclear or if any further question arises — we will respond and address it right away. We are grateful for the rigor of your reading — the paper is more precise, and better, for it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.