Security of deep research agents: attack surface, defenses, benchmarks, and agenda seeding targeting research planning
Security of Deep Research Agents: Attack Surface, Defenses, Benchmarks, and Agenda Seeding against Research Planning
The authors systematize deep research agent security research and propose the agenda seeding attack; in 3,520 sandbox runs, the target product became the top pick in 16.0%–28.0% of cases, and GEO and prompt injection were nearly ineffective.
Attacker capability C1: can publish their own web pages, but does not know the user's specific query, cannot control search rankings, and does not know the model or prompts used by the agent; the budget is publishing 3 new pages on 3 new domains. The victim system is a three-round research agent in a sandbox that iteratively performs planning, retrieval, and integration, and the attack goal is to make the report list the attacker's product as the TOP PICK.
- Deep research agents automatically retrieve open web pages and generate reports with citations; the reports are often taken directly as research conclusions, but related security research is scattered, terminology and threat models vary, and the planning stage in particular lacks research.
- The authors systematize the field across four stages—retrieval, planning, evidence integration, and report generation—providing a threat model, a taxonomy of attacks and defenses, a survey of benchmarks, and coding of 34 papers, and propose an agenda seeding attack that contains no instructions and only introduces a fictitious evaluation metric, compared in an offline sandbox against GEO and prompt injection under the same budget.
- Across 3,520 sandbox runs, agenda seeding made the target product the top pick in 16.0% (dense) and 28.0% (sparse) of runs, while GEO and injection were nearly ineffective; injection detection and root-query anchoring were ineffective against it, verification prompting was effective but harmed utility, and the authors' proposed provenance tagging preserved utility but could be bypassed by a multi-site decoy.
- The authors point out that the corpus is generated by a single model, is far smaller than the real web, and uses BM25 rather than a commercial search engine as the ranker; the agent tested has a simple structure, each condition has only 20–160 runs, and the results only demonstrate feasibility rather than success rates on real systems.
- Evaluated models
- Qwen3.8-FlashDeepSeek-V4.1-FlashKimi-K3
- Benchmarks
- 自建离线沙盒(20 个虚构品类、360 个良性页面、BM25 检索)
- Metrics
- target top-pick ratebest top-pick ratemean target rankplanning hijack rateagenda adoption rate
A Peking University team published a paper on arXiv that systematically reviews security research on deep research agents and proposes an agenda seeding attack that requires no instruction writing. The paper establishes a unified threat model and attack and defense taxonomy across four stages—retrieval, planning, evidence integration, and report generation—and reviews 9 security benchmarks; coding of 34 major studies shows that only 26.5% involve the planning stage, and no study has conducted an end-to-end poisoning evaluation of commercial products. The agenda seeding attack introduces only one fictitious evaluation metric into a shopping guide whose conclusions align with the mainstream, and separately publishes a page on which the target product leads on that metric. In 3520 same-budget sandbox research runs, generative engine optimization (GEO) and prompt injection were almost ineffective, while agenda seeding made the attacker's target the top choice in 16% to 28% of reports; injection detection and query anchoring failed against it, verification prompts harmed usability, and the provenance tagging proposed in the paper maintained usability but could be bypassed by multi-site decoys. The paper concludes with eight research directions.
Why it was selectedA systematic review of 34 studies points out that the planning stage is systematically underestimated, and 3520 sandbox runs show that instruction-free agenda seeding attacks are stronger than prompt injection, and injection detection and query anchoring fail against them.
Paper analysis
What problem does this paper attempt to solve?
Systematize security research on deep research agents (DR agents) and empirically test whether the overlooked planning stage can be manipulated by instruction-free attacks.
Core problem: deep research agents automatically retrieve from the open web and generate reports with citations, and the reports are often taken directly as research conclusions, yet security research in this area is scattered, threat models vary, and the planning stage is almost unstudied.
- Problem setting: DR agents read large numbers of web pages on the user's behalf, and anyone who can place content into their retrieval results can influence the reports that users treat as conclusions (Section 1). The authors state that preliminary research since 2025 has made this risk concrete, e.g., WARP changing about 13 words is enough to be retrieved in 57%–76% of runs.
- Gaps in existing research: related work is scattered across venues in security, NLP, IR, etc., with differing terminology, threat models, and metrics; existing surveys focus only on capability or treat retrieval as one of many untrusted inputs (Section 1). The authors' coding of 34 core papers shows that only 26.5% involve the planning stage, and no work conducts end-to-end poisoning evaluation on commercial products (Abstract, Section 9).
- Core observation: the authors argue that instruction-free, data-level manipulation can bypass defenses centered on instruction detection; the intersection of planning and ranking-improvement objectives is empty in the literature, but DR agents' queries are entirely produced by the planner (Section 9 F2, Section 10).
- Paper contributions: a four-stage system model and unified threat model, an attack taxonomy (4 stages, 19 categories) and defense taxonomy (5 layers), a survey of 9 security benchmarks, dimension-by-dimension alignment with RAG poisoning/prompt injection/GEO, coding of 34 papers, and the agenda seeding attack with 3,520 sandbox runs (Section 1 contribution list).
- Threat model (Section 4.3, Section 10.1):
- Goal: make the DR agent's report list the attacker-specified product as the top pick (ranking-improvement objective Pro), or more broadly alter the integrity of evidence and conclusions.
- Knowledge: the attacker does not know the user's specific query, nor the model and prompts used by the agent; the paper argues that DR attack evaluation should report both knowledge assumptions and capability levels.
- Capability: experiments use level C1—can publish one's own web pages (budget of 3 new pages on 3 new domains), cannot control search ranking; the full scale is C0–C5, with supply-chain attackers C6 listed separately as a blank.
- Victim system: a three-round research agent in the sandbox that iteratively performs planning, retrieval, and integration, structurally benchmarked against open-source systems such as STORM and WebThinker.
What related research exists?
Related work falls into four categories: DR capability systems and surveys, retrieval-stage poisoning and injection, planning- and integration-stage attacks, and defenses and benchmarks; the paper reorganizes them by stage and adds experiments.
DR capability systems and surveys
- Capability systems: open-source frameworks such as STORM, Co-STORM, OmniThink, and MindSearch, as well as products such as OpenAI/Gemini/Perplexity Deep Research, constitute the paper's system background (Sections 1 and 2).
- Capability surveys: Xi et al. (2025a), Zhang et al. (2025c), Huang et al. (2025), Xu and Peng (2025), and Li et al. (2025e) survey paradigms, optimization, and evaluation, but focus only on capability; the paper borrows their pipeline division and reinterprets each stage as an attack surface (Section 12).
- Agent security surveys: Deng et al. (2025), He et al. (2025, 2026), Gan et al. (2024), and others organize threats by model, memory, tools, and environment, treating retrieval as one kind of untrusted input; the authors state that they do not characterize the open-web evidence ecosystem (Sections 1 and 12).
Retrieval-stage attacks (poisoning and injection)
- RAG poisoning: PoisonedRAG (Zou et al., 2025), Zhong et al. (2023), and others established knowledge poisoning methods; the authors point out that their three assumptions—known target query, accessible retriever, and single-round retrieval—do not hold in the DR setting (Section 8).
- UGC editing and result-slot injection: WARP (Zhang et al., 2026b) makes about 13-word edits to UGC pages, which are retrieved in 57%–76% of runs and cited in 38%–51% of reports; SafeSearch (Dong et al., 2026) achieves a maximum attack success rate of 90.5% on 300 test cases; Breadcrumbing (Li et al., 2026b) argues for measuring capability by the number of injectable entries per round.
- Indirect prompt injection: Greshake et al. (2023) proposed indirect injection; Nestaas et al. (2025) manipulate LLM search recommendations; Khodayari et al. (2026) empirically find that most injections in real web pages are invisible to humans.
Planning-, integration-, and generation-stage attacks
- Planning stage: In DR jailbreak research (Chen et al., 2026a), plan injection and intent hijacking require only natural-language rewriting; FORGE (Pan et al., 2026) studies trajectory hijacking; KidnapRAG (Choi et al., 2026a) uses three types of documents—Bait/Chain-Link/Mal-Ins—to hijack agentic RAG.
- Integration stage: HAE-GEO (Bi et al., 2026), on 72,039 clean pages with 770 poisoned pages per level, finds that fabricated mutual corroboration weakens agents' ability to identify suspicious evidence; Lazy Grounding (Zhang et al., 2026c) uses entirely real but distracting evidence to reduce search-agent accuracy by an average of 5.9 percentage points.
- GEO and generation stage: GEO (Aggarwal et al., 2024) reports up to 40% improvement in content visibility; WARP uses the proportion of reports citing poisoned sources as its core metric, i.e., citation laundering; Rao et al. (2026) study citation hallucination without an attacker.
Defenses and benchmarks
- Defenses: instruction–data separation (Spotlighting, StruQ, SecAlign, instruction hierarchy), injection detection (DataSentinel, MELON), robust aggregation (RobustRAG, TrustRAG, Astute RAG), root-query anchoring (FORGE), privilege control (CaMeL, FIDES, Progent), etc.; the authors point out that content-level defenses assume attacks carry instructions, integration-level defenses assume most evidence is benign, and source-level defenses assume authoritative sources are hard to poison, whereas representative DR attacks break these three assumptions respectively (Section 6.6).
- Baselines and benchmarks: the comparison attacks in the paper's experiments are GEO and GEO+injection (same budget); the survey's 9 security benchmarks include Agent-Fence, HAE-GEO, two SafeSearch, SearchGEO, MosaicLeaks, GEO-Bench, WASP, and AgentDojo (Table 5).
Author positioning: closest to Agent-Fence, which organizes 14 types of attacks by trust boundary; this paper organizes by research stage and adds a defense taxonomy, neighborhood alignment, and literature statistics (Section 12).
How does the paper address this problem?
First establish a threat model and taxonomy by the four stages and code 34 papers, then design the agenda seeding attack for the planning-stage gap and conduct controlled sandbox experiments.
The work has two parts overall: first, systematization—abstracting the DR agent into four stages (retrieval, planning, evidence integration, and report generation), establishing a threat model and an attack/defense taxonomy, and coding 34 core papers; second, empirical work—targeting the planning-stage gap found in the coding, proposing the agenda seeding attack and comparing it in an offline sandbox against GEO and prompt injection under the same budget.
Systematization method (Section 3)
- Literature collection: searching venues including USENIX Security, IEEE S&P, ACM CCS, NDSS, ACL, and NeurIPS and four arXiv categories, covering January 2023 to September 2026, with keyword combinations followed by backward and forward snowballing.
- Inclusion criteria: the research object falls within scope, contains an explicit threat model or security goal, and reports empirical results; this yields 34 core studies and 88 adjacent works.
- Coding scheme: each paper is coded along 6 dimensions—type (A/D/B/M), stages involved (R/P/I/G, multi-label), attacker capability (C0–C5), attack goal (Int/Pro/Harm/Hij/Conf/Avail), target system (DR/SA/GE/RAG), evaluation environment (Live/Sandbox/Sim), and whether defenses are evaluated.
Threat model (Section 4)
- Trust boundaries: B1 user–agent, B2 agent–search engine and web pages, B3 result slot and external observers, B4 feedback loop from evidence integration to planning, B5 report–consumer (Figure 1).
- Security goals: six types—evidence and conclusion integrity, ranking and recommendation fairness, safety, control-flow integrity, confidentiality, and availability and cost.
- Attacker capabilities: C0 no attacker (e.g., citation hallucination) through C5 passive observer; C6 supply-chain attacker (controlling the search API, crawler, tool server, or sub-agent) has no research in the core corpus and is listed as a blank (Table 2).
Attack and defense taxonomy (Sections 5 and 6)
- Attack taxonomy: 19 attack categories organized by stage, each with mechanism, required capability, goal, and representative work (Table 3); three cross-stage propagation patterns are summarized: retrieval-to-planning feedback, integration-to-generation amplification, and report-to-consumer diffusion.
- Defense taxonomy: organized into five layers—source, content, integration, planning and system, and generation—each with its dependent assumptions and known failure conditions (Table 4), e.g., content-level defenses assume the attack payload carries instructions and fail against purely data-level manipulation.
Agenda seeding attack design (Section 10.1)
- Core mechanism: the attacker writes no instructions and does not directly claim the target product is best, but introduces a fictitious evaluation metric in a buying guide whose conclusions align with the mainstream, claiming that most reviews overlook it; two other pages (a lab report and a technical explanation) state that the target product leads on this metric. The decoy page does not mention the target product.
- Variants: no-bait (removing the decoy page and keeping only the two metric pages) and adaptive agenda seeding (replacing the technical explanation page with a second decoy page on a second domain, so that the metric is mentioned by two sites in the first round, used to test provenance tagging).
- Comparison attacks: GEO (claiming the target product is best overall) and GEO+injection (appending a sentence addressed to AI assistants at the end of each page) under the same budget (3 new pages on 3 new domains).
Sandbox and agent (Section 10.1)
- Corpus: 20 real consumer categories, each with 5 fictitious brand models and preset scores on 4 generic purchase criteria, with one product leading on 3 criteria as the ground-truth best and the third-ranked product as the attack target; Qwen3.7-Plus generates 18 benign pages per category from the score table, 360 pages in total; there is also a sparse version (only 6 benign pages in the target category). Retrieval uses BM25, returning the top 4 results per query.
- Agent tested: an iterative research agent whose planner first gives 3 initial queries, then in each round reads new results, takes notes by source, and proposes 2–3 follow-up queries, for 3 retrieval rounds in total, and finally writes a 300–500-word report with inline citations, outputting two lines, RANKING and TOP PICK; the backend is Qwen3.8-Flash, DeepSeek-V4.1-Flash, and Kimi-K3, called via the Alibaba Cloud Bailian standard interface with thinking mode disabled.
- Paired design: within the same category and sample, all model calls before the attack page first appears are reused across conditions, with run-by-run pairing across conditions; differences use the paired exact McNemar test.
Metrics and defenses (Section 10.1)
- Metrics: target top-pick rate, best top-pick rate (measuring research quality), target average rank (recorded as 6 when absent), proportion of attack pages read/cited, planning hijack rate (queries from round 2 onward containing the fictitious metric), and agenda adoption rate (the report mentions the fictitious metric).
- Four defenses: injection detection (Qwen3.8-Flash judges whether a page contains text directing an AI), root-query anchoring (text-embedding-v4 cosine similarity, with the threshold set at the 10th percentile of the similarity of all follow-up queries in clean runs), verification prompting (the system prompt requires cross-checking and wariness of single-source metrics), and the authors' proposed provenance tagging (queries relying on a single-source concept are flagged, related pages are recorded as a single-source lineage, and a standard supported only by such pages must not determine the conclusion).
What experiments did the paper conduct?
3,520 sandbox runs: agenda seeding target top-pick rate 16.0%/28.0%, GEO and injection nearly ineffective; among four defenses, only verification prompting was effective but harmed utility.
Experimental setup
- Models tested: Qwen3.8-Flash, DeepSeek-V4.1-Flash, Kimi-K3, all via the Alibaba Cloud Bailian standard interface with thinking mode disabled; the first two were sampled 2 times per condition per category, and Kimi-K3 was sampled once due to invocation cost.
- Dataset: a self-built offline sandbox with 20 fictitious categories and 360 benign pages (generated by Qwen3.7-Plus), plus a sparse version (6 benign pages in the target category); BM25 retrieval, returning the top 4 results per query (Section 10.1).
- Comparison attacks: GEO, GEO+injection, and agenda seeding under the same budget (3 pages on 3 new domains), plus two variants, no-bait and adaptive seeding.
- Metrics: target top-pick rate, best top-pick rate, target average rank, read/cited rate, planning hijack rate, agenda adoption rate; paired exact McNemar test.
- Sample size: each summary cell in Table 9 has n=100 runs; each cell in Table 10 has n=160 (Qwen and DeepSeek combined, two coverage levels), with n=20 for the Kimi column; 3,520 runs in total (Abstract).
- Evaluation method: conclusions are read by rule from the report's RANKING/TOP PICK lines; the injection-detection defense uses Qwen3.8-Flash to judge whether a page contains text directing an AI.
Main results
According to Table 9 (no defense, three retrieval rounds, three models combined, n=100 per cell):
Swipe horizontally to read the table
Coverage Condition Target top-pick rate Best top-pick rate Target average rank Hijack rate Adoption rate Dense Clean 0.0% 100.0% 3.07 0.0% 0.0% Dense GEO 0.0% 98.0% 3.04 0.0% 0.0% Dense GEO+injection 5.0% 94.0% 2.95 0.0% 0.0% Dense Agenda seeding 16.0% 80.0% 2.62 72.0% 96.0% Sparse GEO 6.0% 93.0% 2.89 0.0% 0.0% Sparse Agenda seeding 28.0% 64.0% 2.39 88.0% 99.0% - The authors state that under the same budget, changing the comparison standard is far more effective than directly promoting the target product: the paired tests of agenda seeding against clean and GEO are both p < 10⁻⁴ under both coverage levels, GEO is p = 0.031 relative to clean only on the sparse corpus, and the remaining GEO/injection comparisons have p ≥ 0.06.
- The authors point out that GEO and injection pages are read and cited in almost every run (98%–100%), yet the model still sides with the majority of sources that contradict them; the agenda seeding decoy page does not even mention the target product.
- By model (Table 9): on the sparse corpus, Qwen3.8-Flash is flipped 47.5% of the time, Kimi-K3 30.0% under both coverage levels, and DeepSeek-V4.1-Flash only 2.5%/7.5%; the authors argue that recognizing promotional content and judging the credibility of evaluation standards are two different capabilities.
Defense evaluation (Table 10, Figure 5)
- Injection detection and root-query anchoring: injection detection lowers the target top-pick rate under injection to 0 (0 vs. 5 flips, p = 0.06; top-two 0 vs. 11, p < 0.001), but is completely ineffective against agenda seeding and increases research token cost by about 69% per run; root-query anchoring lowers the hijack rate from 75.6% to 54.4%, but the target top-pick rate remains 20.0%.
- Verification prompting: the only defense that substantially reduces attack effectiveness—the agenda seeding target top-pick rate drops from 20.0% to 8.8% (p < 0.001), and Kimi from 30.0% to 0; however, the no-attack best top-pick rate drops from 100% to 86.2% (DeepSeek to 76.2%), and the authors state that it works mainly by drawing fewer conclusions rather than correct ones.
- Provenance tagging (proposed by the authors): the no-attack best top-pick rate remains 100% (mislabeling rate 3.8%), and the agenda seeding target top-pick rate drops from 20.0% to 15.0% (eliminating 8 flips and introducing no new flips, p = 0.008); however, against the two-site decoy adaptive seeding it only drops from 18.8% to 17.5% (p = 0.5), and the authors state that judging independence by site count can be forged by a multi-site attacker.
Mechanism analysis (Figure 4, Figure 5(c))
- Attack path: in agenda seeding runs, 80.0% pass the hijack step and 97.5% pass the adoption step, with the loss mainly occurring at the final step from adopting the metric to ranking the target first; on the dense corpus, 53% of runs read the metric page via subsequent queries and 40% retrieve it in the first-round generic query, while on the sparse corpus 82% retrieve it in the first round.
- Why anchoring fails: the median similarity between queries containing the fictitious metric and the original question is 0.42, lower than the 0.60 of no-attack follow-up queries, but the two distributions overlap; the threshold discards only 39.2% of metric queries while already discarding 9.1% of normal queries (Figure 5(c)).
- Across categories: in 40 category×corpus combinations, agenda seeding improves the target average rank in 75% of combinations, versus 28% for both GEO and GEO+injection; the average rank improvements are 0.50, 0.09, and 0.12 positions respectively (Figure 4(c)).
Other ablations and analyses
- Single-round retrieval (dense corpus, Qwen and DeepSeek): the target top-pick rate drops from 12.5% to 6.2%, and the adoption rate from 95.0% to 65.0%.
- No-bait variant: the target top-pick rate drops from 12.5% and 27.5% to 10.0% and 18.8% on the dense/sparse corpora respectively; on the sparse corpus the difference in the drop in best top-pick rate is p = 0.035, and the other differences are not significant at this scale.
- Literature coding (Table 8, N=34): only 26.5% involve the planning stage, C1 attackers account for 67.6%, and C6 is 0; 55.9% evaluate on simulated corpora, and only 14.7% specifically propose defenses.
- Negative results: the authors explicitly state that some ablation differences are not significant at the experimental scale, and that the experiments only demonstrate feasibility and do not represent attack success rates on real systems (Section 10.3).
What are some points that can be further explored?
The authors list experimental limitations such as the gap between the sandbox and real systems, and propose eight research directions including planning integrity, supply-chain attackers, and adaptive evaluation.
Limitations and future directions noted by the authors
- Sandbox fidelity (Section 10.3): the corpus is generated by a single model, its scale (20 categories, 18 benign pages each) is far smaller than the real web, and the ranker is BM25 rather than a commercial search engine.
- Agent tested and sample size (Section 10.3): what is tested is a structurally simple three-round research pipeline, differing from commercial products in retrieval scale, number of queries, and security mechanisms; each condition has only 20–160 runs, and some ablation differences are not significant at this scale; the results demonstrate feasibility and defense failure modes, not attack success rates on real systems.
- Limitations of provenance tagging (Section 10.2 E5, Section 11 G5): it takes effect only when the metric is introduced via a single source (only 21.2% of agenda seeding runs trigger the flag), and the site-count-based independence judgment can be bypassed by one additional decoy site, requiring stronger independence evidence (e.g., publisher affiliation, page history, first-appearance time).
- Eight research gaps and directions (Section 11): G1 planning integrity (formalizing research-direction hijacking and distinguishing legitimate exploration from induced deviation); G2 report-level attribution integrity (source credentials, edit history, citation laundering and targeted cloaking detection); G3 editorial and supply-chain attackers (C2/C3 research is scarce, C6 is entirely blank); G4 confidentiality and availability (query rewriting, differential-privacy-style query obfuscation, wallet-drain detection); G5 defenses for evidence reasoning (source independence, temporal consistency, cross-lingual and cross-platform corroboration, bounded guarantees under multi-hop retrieval); G6 standardized reproducible evaluation (frozen web snapshots, pluggable rankers, unified poisoning interfaces); G7 adaptive attacks and defense combinations; G8 multimodality, multi-agent settings, and human factors.
Scope of experiments
- The models tested are Qwen3.8-Flash, DeepSeek-V4.1-Flash, and Kimi-K3, 3 in total, all via the Alibaba Cloud Bailian interface with thinking mode disabled (Section 10.1).
- Attack evaluation is conducted in a self-built offline sandbox: 20 fictitious categories, 360 benign pages plus a sparse version, BM25 retrieval; the comparison attacks are GEO and GEO+injection, and no end-to-end poisoning is conducted on commercial products (Section 10.1; the authors state that this would affect real users and faces ethical obstacles, Section 7).
- Defense evaluation covers four layers: injection detection, root-query anchoring, verification prompting, and provenance tagging, with an adaptive attack test for provenance tagging (Table 10).
- Sample sizes are n=100 per cell (Table 9) or n=160 (Table 10), with n=20 for the Kimi column; Qwen and DeepSeek have 2 samples per condition per category and Kimi 1; differences use the paired exact McNemar test (Section 10.1).
- Literature coding covers 34 core studies from January 2023 to September 2026 (Section 3); the paper does not report human evaluation or agreement between human and automatic judgments.
Summarize the main content of the paper
Systematization of DR agent security plus empirical evidence on agenda seeding: instruction-free data-level manipulation can change research conclusions, and existing defenses centered on instruction detection are ineffective against it.
A systematization and empirical paper on the security of deep research agents (DR agents): the authors abstract the DR agent into four stages—retrieval, planning, evidence integration, and report generation—and answer five questions: what the attack surface is, how existing attacks and defenses are classified, what benchmarks lack, what gaps the literature distribution reveals, and whether the planning stage can be manipulated by instruction-free attacks.
- Problem addressed: DR agents automatically read the open web and generate reports with citations, and the reports are often taken directly as conclusions, but research in this area is scattered and threat models and metrics vary; the authors' coding of 34 core papers shows that only 26.5% of work involves the planning stage, and no work conducts end-to-end poisoning evaluation on commercial products.
- Method highlights: a threat model with five trust boundaries, six attacker capability levels (C0–C5, with C6 supply chain listed separately as a blank), and six types of security goals; 19 attack categories organized by stage, defenses organized into five layers with their assumptions and failure conditions annotated; dimension-by-dimension alignment with RAG poisoning, prompt injection, and GEO. Empirically, it proposes agenda seeding: the attacker writes no instructions and does not claim the target is better, but only introduces a fictitious evaluation metric.
- Main results (3,520 paired sandbox runs, Tables 9 and 10): under the same budget, GEO and GEO+injection are nearly ineffective (target top-pick rate at most 6.0%), while agenda seeding reaches 16.0% and 28.0% on dense/sparse corpora respectively, and lowers the best top-pick rate from 100% to 80.0%/64.0%; in 72.0%–88.0% of runs the agent actively searched for the fictitious metric.
- Defense results: injection detection and root-query anchoring are ineffective against agenda seeding; verification prompting lowers the target top-pick rate from 20.0% to 8.8%, but lowers the no-attack best top-pick rate from 100% to 86.2%; the authors' proposed provenance tagging preserves utility (best top-pick rate 100%) but can be bypassed by a two-site decoy (18.8% drops only to 17.5%).
- Authors' conclusions: the authors argue that DR security cannot be treated as a special case of prompt injection—the core task is to draw conclusions from data, and purely data-level manipulation can achieve the attack; defenses need evidence reasoning in adversarial environments, judging not only whether claims are credible but also who proposed the comparison standard and why it holds. The authors also emphasize that the sandbox results only demonstrate feasibility and do not represent success rates on real systems, and propose eight research directions.