Security of Deep Research Agents: Attack Surface, Defenses, Benchmarks, and Agenda Seeding against Research Planning
Organizations: Peking University
Abstract
Deep research and agentic search agents autonomously retrieve content from the open web and generate reports with citations, and their outputs are often taken directly as research conclusions. This paper systematizes security research in this area along four stages, namely retrieval, planning, evidence integration, and report generation. We present a unified threat model, a taxonomy of attacks and defenses, and a review of evaluation benchmarks, and we align the area dimension by dimension with retrieval-augmented generation (RAG) poisoning, prompt injection, and generative engine optimization (GEO). Coding 34 primary studies shows that only 26.5% of them involve the planning stage, and none performs an end-to-end poisoning evaluation on commercial products. To address this gap, we propose the agenda seeding attack, in which the attacker writes no instructions and only introduces a fictitious evaluation metric. Across 3,520 sandboxed research runs under the same budget, GEO and prompt injection are nearly ineffective, whereas agenda seeding makes the attacker's target the top pick in 16% to 28% of reports. Injection detection and query anchoring fail against it, verification prompting hurts utility, and our proposed provenance tagging preserves utility but can be bypassed by multi-site bait. Finally, we propose eight research directions.
Figures & tables
| Property | RAG | Web agent | Agentic search | Deep research |
|---|---|---|---|---|
| Corpus | Curated or private corpus | Live websites under interaction | Open web via search API | Open web, UGC, PDFs, and private documents |
| Retrieval rounds | One per question | Driven by actions | Several, adaptive | Tens to hundreds, adaptive and recursive |
| Query author | User question | Agent | Agent | Planner and sub-agents |
| Output | Short answer | Actions with side effects | Short answer with citations | Long structured report with citations |
| Dominant harm | Wrong answer | Unsafe action | Wrong answer | Misleading report trusted as a research conclusion |
| Examples | RAG ( Lewis et al., 2020 ) | WebArena ( Zhou et al., 2024 ) | Search-R1 ( Jin et al., 2025 ) | STORM ( Shao et al., 2024 ) , OpenAI DR ( OpenAI, 2025 ) |
| Level | Capability | Instances and representative works |
|---|---|---|
| C0 | No adversary | Fabricated or mismatched references ( Rao et al., 2026 ) |
| C1 | Publish own pages | Search-optimized and GEO pages, content farms ( Aggarwal et al., 2024 ; Bi et al., 2026 ; Zhang et al., 2026c ) |
| C2 | Edit existing third-party pages | UGC comments, wiki or review edits ( Zhang et al., 2026b ; Caville et al., 2026 ) |
| C3 | Control result slots | Injected search results, ads, malicious search tools ( Dong et al., 2026 ; Li et al., 2026b ) |
| C4 | Submit queries | Jailbreak and intent hijacking ( Chen et al., 2026a ; Zhan et al., 2026 ) |
| C5 | Observe queries or traffic | Search providers, network observers ( Gurung et al., 2026 ; Jeong et al., 2026 ) |
| Stage | Attack class | Mechanism | Cap. | Goal | Representative works |
|---|---|---|---|---|---|
| R | Open-web corpus poisoning | Publish pages crafted to be retrieved for target topics | C1 | Int, Pro | ( Bi et al., 2026 ; Pan et al., 2023 ; Zou et al., 2025 ; Zhong et al., 2023 ) |
| UGC and third-party edit poisoning | Small edits to pages the agent already retrieves often | C2 | Pro, Int | ( Zhang et al., 2026b ; Caville et al., 2026 ) | |
| Result-slot injection | Insert unreliable results into the result list, possibly per turn | C3 | Int, Harm | ( Dong et al., 2026 ; Li et al., 2026b ) | |
| Retrieval-borne prompt injection | Instructions hidden in retrieved pages | C1 | Hij, Conf | ( Nestaas et al., 2025 ; Khodayari et al., 2026 ; Chang et al., 2026 ; Rall et al., 2025 ) | |
| Agent-targeted cloaking | Serve different content to agent crawlers and to humans | C1 | Int, Hij | ( Wang et al., 2011 ) | |
| Query and traffic leakage | Infer private context from outbound queries or metadata | C5 | Conf | ( Gurung et al., 2026 ; Jeong et al., 2026 ) |
| Layer | Defense class | Representative works | Assumption | Known failure condition |
| Source | Credibility and authority scoring | ( Jia et al., 2025 ; Mochizuki et al., 2025 ) | Authority correlates with integrity | Edits to high-authority or UGC pages ( Zhang et al., 2026b ; Caville et al., 2026 ) |
| Source exclusion | ( Zhang et al., 2026b ) | Risky source types can be enumerated | Coverage loss, and poisoned non-UGC pages remain | |
| Content | Instruction and data separation | ( Hines et al., 2024 ; Chen et al., 2025a ; Chen et al., 2025b ; Wallace et al., 2024 ) | Attack payload carries instructions | Data-only manipulation ( Zhang et al., 2026b ; Zhang et al., 2026c ) and adaptive attacks ( Zhan et al., 2025 ; Nasr et al., 2026 ) |
| Injection and anomaly detection | ( Liu et al., 2025 ; Zhu et al., 2025 ; Luo et al., 2025 ; Yu et al., 2026 ) | Manipulation is anomalous in form | Semantic and persuasion attacks evade filters ( Yu et al., 2026 ; Zhang et al., 2026b ) | |
| Integration | Robust aggregation and certification | ( Xiang et al., 2024 ; Zhou et al., 2025 ; Wei et al., 2025b ; Wang et al., 2025 ) | Most retrieved evidence is benign | Coordinated corroboration ( Bi et al., 2026 ; Li et al., 2026b ) , and multi-hop retrieval is not covered |
| Verification prompting | ( Dong et al., 2026 ; Bi et al., 2026 ; Chen et al., 2026b ) | Model follows verification instructions | Verification rarely becomes recovery ( Bi et al., 2026 ) |
| Benchmark | Target | Stages | Threat | Environment | Main metrics |
|---|---|---|---|---|---|
| Agent-Fence ( Puppala et al., 2026 ) | 8 deep agent archetypes | R, P, I, G | 14 trust-boundary attack classes | Controlled, fixed-model protocol | Trace-auditable breaks, per-class risk |
| HAE-GEO ( Bi et al., 2026 ) | 10 deep search agents | R, I | 3-level evidence poisoning | 72,039 clean and 770 poisoned pages per level | Recognition, verification, recovery, resistance |
| SafeSearch ( Dong et al., 2026 ) | 3 scaffolds, many backends | R, I | 5 risk categories, 300 cases | Real results plus one injected site | Checklist-based attack success |
| SafeSearch ( Zhan et al., 2026 ) | Search agents trained with reinforcement learning | P, G | Harmful queries | Live search | Harmfulness and utility |
| SearchGEO ( Chen et al., 2026b ) | 13 search backends | I | Attacker-published content | Controlled web content | Endorsement of manipulated content |
| MosaicLeaks ( Gurung et al., 2026 ) | Research agents | P, R | Query-log adversary | 1,001 tasks over private and public corpora | Leakage of intent, answers, and claims |
| Dimension | RAG poisoning | Prompt injection | GEO and search engine manipulation | Deep research and agentic search |
|---|---|---|---|---|
| Injection point | Curated or private corpus | Any untrusted input | Public web content and ranking | Open web, result slots, UGC, memory |
| Knowledge of query | Usually known | Not required | Topic level | Unknown, and the agent rewrites queries |
| Payload | Adversarial passage | Instruction | Persuasive or optimized content | Data, instruction, or both |
| Retrieval rounds | One | Per tool call | One ranking | Many, adaptive, recursive |
| Attack goal | Targeted wrong answer | Hijack task or tools | Visibility and rank | Conclusions, citations, plans, privacy, cost |
| Transferable defense | Robust aggregation | Instruction and data separation | Spam and cloaking detection | Partial, see Section 6 |
| Work | Year | Type | R | P | I | G | Attacker | Goal | Target | Env. | Def. |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Agent-Fence ( Puppala et al., 2026 ) | 2026 | B | ✓ | ✓ | ✓ | ✓ | C1, C4 | Int, Hij, Conf, Avail | DR | Sandbox | |
| WARP ( Zhang et al., 2026b ) | 2026 | A | ✓ | ✓ | ✓ | C2 | Pro, Int | DR | Sim | Y | |
| HAE-GEO ( Bi et al., 2026 ) | 2026 | B | ✓ | ✓ | C1 | Int | SA | Sim | Y | ||
| DR Jailbreak ( Chen et al., 2026a ) | 2025 | A | ✓ | ✓ | C4 | Harm | DR | Live | |||
| DR Guardrails ( Huang et al., 2026 ) | 2025 | D | ✓ | ✓ | ✓ | C1, C4 | Harm, Int | DR | Live | Y | |
| FORGE ( Pan et al., 2026 ) | 2026 | A | ✓ | C1 | Int, Hij | DR | Sim | Y |
| Stage | Int | Pro | Harm | Hij | Conf | Avail | Papers |
| R | 11 | 3 | 3 | 7 | 4 | 1 | 17 (50.0%) |
| P | 5 | 0 | 4 | 3 | 2 | 1 | 9 (26.5%) |
| I | 9 | 12 | 1 | 3 | 1 | 1 | 19 (55.9%) |
| G | 7 | 8 | 5 | 2 | 1 | 1 | 17 (50.0%) |
| All | 17 | 12 | 6 | 8 | 4 | 1 | 34 |
| Attacker: C1 23, C4 5, C2 3, C3 2, C5 2, C0 1, C6 0 | |||||||
| Target top pick (%) | Pooled over the three models | ||||||||||
| Coverage | Condition | Qwen | DeepSeek | Kimi | Target (%) | Best (%) | Target rank | Read (%) | Cited (%) | Hijack (%) | Adopt (%) |
| Dense | Clean | 0.0 | 0.0 | 0.0 | 0.0 | 100.0 | 3.07 | n/a | n/a | 0.0 | 0.0 |
| GEO | 0.0 | 0.0 | 0.0 | 0.0 | 98.0 | 3.04 | 100.0 | 100.0 | 0.0 | 0.0 | |
| GEO + injection | 10.0 | 0.0 | 5.0 | 5.0 | 94.0 | 2.95 | 100.0 | 98.0 | 0.0 | 0.0 | |
| Agenda seeding | 22.5 | 2.5 | 30.0 | 16.0 | 80.0 | 2.62 | 99.0 | 97.0 | 72.0 | 96.0 | |
| Sparse | Clean | 0.0 | 0.0 | 0.0 | 0.0 | 100.0 | 3.02 | n/a | n/a | 0.0 | 0.0 |
| Clean | Injection | Agenda seeding | Adaptive seeding | Tokens per clean run | Kimi seeding | ||||
|---|---|---|---|---|---|---|---|---|---|
| Layer | Defense | Best (%) | Target (%) | Target (%) | Best (%) | Hijack (%) | Target (%) | Target (%) | |
| None | 100.0 | 3.1 | 20.0 | 73.1 | 75.6 | 18.8 | 10.4k | 30.0 | |
| Content | Injection detection | 100.0 | 0.0 | 20.0 | 73.1 | 75.6 | n/a | 17.6k | 30.0 |
| Planning | Root-query anchoring | 100.0 | 1.9 | 20.0 | 75.0 | 54.4 | n/a | 10.3k | 30.0 |
| Integration | Verification prompting | 86.2 | 1.2 | 8.8 | 66.2 | 56.9 | 6.9 | 10.7k | 0.0 |
| Lineage | Provenance tagging (ours) | 100.0 | 3.1 | 15.0 | 78.1 | 74.4 | 17.5 | 10.5k | 25.0 |
| Survey | Focus | Pipe. | Threat | Atk. | Def. | Adj. | Coding | Exp. |
|---|---|---|---|---|---|---|---|---|
| Xi et al. (2025a) | Deep search agents, capability | ✓ | ||||||
| Zhang et al. (2025c) | Deep research agents, capability | ✓ | ||||||
| Huang et al. (2025) | Deep research agents, capability | ✓ | ||||||
| Xu and Peng (2025) | Deep research systems, capability | ✓ | ||||||
| Deng et al. (2025) | LLM agent security | ✓ | ✓ | |||||
| He et al. (2026) | LLM agent security and privacy | ✓ | ✓ |
Deep Research Agents
Deep research agents plan long-horizon searches, synthesize cited evidence, and produce reports for complex open-ended questions.
See all 120 papers