03 arXiv 論文精選(67 篇,連結全部於 2026-10-26 以 arxiv.org/abs 驗證標題)
索引:README|期刊/廠商見 04-journals-industry|總表 references
說明:作者欄列前 1–3 位,et al. 省略。2026 年論文多為 preprint,未經同儕審查,標 ⚠️。takeaway 依摘要撰寫。
A. 奠基:prompt injection 的發現與形式化
- Ignore Previous Prompt: Attack Techniques For Language Models — Perez & Ribeiro, 2022. https://arxiv.org/abs/2211.09527 早期系統化 goal hijacking 與 prompt leaking(PromptInject)。關聯:歷史起點。
- Not what you’ve signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection — Greshake et al., 2023. https://arxiv.org/abs/2302.12173 首次定義 indirect prompt injection,在 Bing Chat 等實際系統示範資料竊取、蠕蟲式傳播、遠端控制。關聯:所有 agent 資安的祖師論文(AISec ‘23)。
- Prompt Injection attack against LLM-integrated Applications — Liu et al., 2023. https://arxiv.org/abs/2306.05499 HouYi 攻擊,36 個真實應用中 31 個可被注入。關聯:說明「應用層」普遍脆弱。
- Formalizing and Benchmarking Prompt Injection Attacks and Defenses — Liu, Jia, Gong et al., 2023. https://arxiv.org/abs/2310.12815 提出形式化框架與組合攻擊,評估多種防禦(USENIX Security ‘24)。關聯:理解「防禦評估」的基本方法。
- Automatic and Universal Prompt Injection Attacks against LLMs — Liu et al., 2024. https://arxiv.org/abs/2403.04957 以梯度最佳化自動產生通用注入。關聯:攻擊可自動化、可規模化。
B. Jailbreak(與 injection 的區別)
- Jailbroken: How Does LLM Safety Training Fail? — Wei, Haghtalab, Steinhardt, 2023. https://arxiv.org/abs/2307.02483 提出 competing objectives 與 mismatched generalization 兩個失效機制。關聯:解釋為何 safety training 本質上可繞過。
- Universal and Transferable Adversarial Attacks on Aligned Language Models (GCG) — Zou et al., 2023. https://arxiv.org/abs/2307.15043 自動 suffix 攻擊可跨模型轉移,包含閉源模型。關聯:閉源≠安全。
C. Agent 基準測試(benchmarks)
- InjecAgent — Zhan et al., 2024. https://arxiv.org/abs/2403.02691 1,054 個 IPI 測例、17 種使用者工具、62 種攻擊者工具;ReAct GPT-4 有相當比例被攻破。關聯:工具型 agent 的第一個大規模 IPI benchmark(ACL Findings ‘24)。
- AgentDojo — Debenedetti et al., 2024. https://arxiv.org/abs/2406.13352 動態環境(workspace、banking、travel、slack),同時量 utility 與 ASR;已成業界標準(NeurIPS ‘24 D&B)。關聯:評估 某醫學中心 工具鏈可參考其設計。
- Agent Security Bench (ASB) — Zhang et al., 2024. https://arxiv.org/abs/2410.02644 10 情境、400+ 工具、多種攻擊(DPI、OPI、memory poisoning、PoT backdoor)與防禦,平均 ASR 高(ICLR ‘25)。
- Identifying the Risks of LM Agents with an LM-Emulated Sandbox (ToolEmu) — Ruan et al., 2023. https://arxiv.org/abs/2309.15817 用 LLM 模擬工具執行以低成本找出高風險行為。關聯:上線前風險篩檢。
- R-Judge — Yuan et al., 2024. https://arxiv.org/abs/2401.10019 評估 LLM 能否判斷 agent 軌跡的安全風險。關聯:LLM-as-monitor 的上限。
- AgentHarm — Andriushchenko et al., 2024. https://arxiv.org/abs/2410.09024 衡量 agent 執行惡意多步任務的意願;簡單 jailbreak 即可讓 agent 保持能力完成惡意任務。
- WASP: Benchmarking Web Agent Security Against Prompt Injection — Evtimov et al., 2025. https://arxiv.org/abs/2504.18575 真實 web 環境;agent 常「開始照做」但未必完成——「security by incompetence」。
- Benchmarking and Defending Against Indirect Prompt Injection Attacks on LLMs (BIPIA) — Yi et al., 2023. https://arxiv.org/abs/2312.14197 BIPIA benchmark,並提出 black-box(boundary/示範)與 white-box(對抗訓練)防禦;指出 LLM 無法區分指令與外部內容是主因。
- MCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP Servers — 2025. https://arxiv.org/abs/2508.14925 在真實 MCP server 上評測 tool poisoning;能力越強的模型越容易照描述做。關聯:直接對應 某醫學中心 MCP gateway。
D. 防禦:訓練/提示層(機率性)
- Spotlighting — Hines et al. (Microsoft), 2024. https://arxiv.org/abs/2403.14720 delimiting / datamarking / encoding 標記不可信資料,大幅降低 ASR。關聯:成本低,但非保證。
- StruQ — Chen et al., 2024. https://arxiv.org/abs/2402.06363 把 prompt 與 data 分到不同通道並以特殊格式微調(USENIX Security ‘25)。
- SecAlign — Chen et al., 2024. https://arxiv.org/abs/2410.05451 用 preference optimization 讓模型偏好忽略資料中的指令(CCS ‘25)。
- The Instruction Hierarchy — Wallace et al. (OpenAI), 2024. https://arxiv.org/abs/2404.13208 訓練模型依 system > user > tool 優先順序。關聯:閉源模型主要依賴的軟邊界。
- Jatmo — Piet et al., 2023. https://arxiv.org/abs/2312.17673 任務專用微調的非指令模型天然免疫多數注入。關聯:單一任務的醫療抽取器可考慮。
- MELON — Zhu et al., 2025. https://arxiv.org/abs/2502.05174 遮蔽使用者指令重跑一次,若工具呼叫相同即判定被注入(ICML ‘25)。
E. 防禦:系統/架構層(可證明或決定性)
- Defeating Prompt Injections by Design (CaMeL) — Debenedetti et al. (Google DeepMind), 2025. https://arxiv.org/abs/2503.18813 Privileged LLM 只看使用者請求產生程式;Quarantined LLM 處理不可信資料但不能呼叫工具;值帶 capability,由 interpreter 強制資料流政策。AgentDojo 上以可證明方式解決多數任務。關聯:最接近「架構優先」防禦觀點的 SOTA。
- Design Patterns for Securing LLM Agents against Prompt Injections — Beurer-Kellner et al., 2025. https://arxiv.org/abs/2506.08837 六種模式:Action-Selector、Plan-Then-Execute、LLM Map-Reduce、Dual LLM、Code-Then-Execute、Context-Minimization,並以 10 個案例(含醫療問診摘要)說明取捨。關聯:設計 某醫學中心 工具流程的工具書。
- Securing AI Agents with Information-Flow Control (Fides) — Costa, Köpf et al. (Microsoft), 2025. https://arxiv.org/abs/2505.23643 形式化 planner 的安全性,追蹤 confidentiality/integrity label,決定性地強制政策。
- Progent: Securing AI Agents with Privilege Control — Shi et al., 2025. https://arxiv.org/abs/2504.11703 以 DSL 描述工具呼叫權限政策,AgentDojo ASR 降到接近 0 且保留 utility。關聯:MCP gateway 層 policy engine。
- ⚠️ SPIDES: Information-Flow Control for Browser Agents — Zverev et al., 2026. https://arxiv.org/abs/2610.24277 主張恢復 SOP 既不充分也不必要,改以 IFC gate 瀏覽器原語動作。
- ⚠️ Kairos: Transactional Security Enforcement at the System Boundary for LLM Agents — Hempel et al., 2026. https://arxiv.org/abs/2610.18457 單一動作看不出攻擊,需在系統邊界以「交易」為單位中介 agent 的副作用。
- ⚠️ Agentic AI Defense: Decode-Time Provenance Binding for MCP — Yocam et al., 2026. https://arxiv.org/abs/2610.31974 圖片隱寫指令經工具解碼進入 MCP 流程(cross-modal trust laundering);以 HMAC 信任標籤+provenance ledger+taint tracking 阻擋。關聯:MCP gateway 可直接借鏡。
- ⚠️ Prismata: Confining Cross-Site Prompt Injection in Web Agents — 2026. https://arxiv.org/abs/2607.08147
F. 打穿防禦(adaptive attacks)
- Adaptive Attacks Break Defenses Against IPI on LLM Agents — Zhan et al., 2025. https://arxiv.org/abs/2503.00061 8 種防禦全部被 adaptive attack 繞過(NAACL Findings ‘25)。
- The Attacker Moves Second — Nasr, Carlini et al., 2025. https://arxiv.org/abs/2510.09023 對 12 種近期 jailbreak/PI 防禦做 adaptive(梯度、RL、搜尋、人類紅隊)攻擊,多數 ASR >90%。結論:靜態 benchmark 的低 ASR 不代表安全。
- ⚠️ From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents — Yang et al., 2026. https://arxiv.org/abs/2610.22987 IPI 可操弄 agent 提出的 allow-rule,讓使用者核准過寬權限。關聯:HITL 本身是攻擊面。
- ⚠️ Ignored, Not Refused — Zarif et al., 2026. https://arxiv.org/abs/2610.23254 小模型 ASR 低多半是「沒看懂」而非「拒絕」;只改 system prompt ASR 0.2%→43%。關聯:反駁「憑體感選模型」。
G. Memory / RAG / 多 agent
- AgentPoison — Chen et al., 2024. https://arxiv.org/abs/2407.12784 以 <0.1% 投毒率對 memory/RAG 植入 backdoor,平均攻擊成功率高且幾乎不影響正常表現(NeurIPS ‘24)。
- ⚠️ After the Attacker Leaves: Carryover Attacks (CarryoverBench) — Xu et al., 2026. https://arxiv.org/abs/2610.28425 一次注入寫入 memory 檔或排程任務後可跨 session 重複觸發。
- Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems — Lee & Tiwari, 2024. https://arxiv.org/abs/2410.07283
- Agent Smith — Gu et al., 2024. https://arxiv.org/abs/2402.08567 單一對抗圖片在多模態 agent 群中指數級感染(ICML ‘24)。
- Here Comes The AI Worm (Morris II) — Cohen, Bitton, Nassi, 2024. https://arxiv.org/abs/2403.02817 self-replicating prompt 透過 RAG email 助理零點擊傳播。
- ⚠️ The Model Knows Better: Defending Medical Multi-Agent Systems Against Prompt Injection — Sakib, Zhu, Uzuner, 2026. https://arxiv.org/abs/2610.22594 醫療多 agent 討論系統中單一注入段落成功率最高 61%;協作不能保護;看起來像臨床文字的注入最難擋;讓模型對照病人 findings 檢查可大幅回復。關聯:與臨床場景最直接相關。
H. MCP 專屬
- Model Context Protocol (MCP): Landscape, Security Threats, and Future Research Directions — Hou et al., 2025. https://arxiv.org/abs/2503.23278 MCP server 生命週期(creation/operation/update)的威脅分類。
- MCP Safety Audit — Radosevich & Halloran, 2025. https://arxiv.org/abs/2504.03767 示範 Claude/Llama 透過 MCP 被誘導執行 malicious code、remote access、credential theft;提出 MCPSafetyScanner。
- ⚠️ Structured Authority Injection: Persistent Human Control via Typed MCP Overrides — 2026. https://arxiv.org/abs/2610.22373
I. Web / Computer-use agent
- EIA: Environmental Injection Attack on Generalist Web Agents — Liao et al., 2024. https://arxiv.org/abs/2409.11295
- Attacking Vision-Language Computer Agents via Pop-ups — Zhang, Yu, Yang, 2024. https://arxiv.org/abs/2411.02391 簡單彈窗即讓 computer-use agent 點擊,平均 ASR 高達 86%。關聯:支持「computer use 被過度強調」——GUI 是最大、最難控的攻擊面。
- Commercial LLM Agents Are Already Vulnerable to Simple Yet Dangerous Attacks — Li et al., 2025. https://arxiv.org/abs/2502.08586 無需 ML 知識,透過 Reddit 貼文等即可誘導商用 agent 洩漏信用卡、下載惡意檔、寄釣魚信。
- Why Are Web AI Agents More Vulnerable Than Standalone LLMs? — 2025. https://arxiv.org/abs/2502.20383
- Simple Prompt Injection Attacks Can Leak Personal Data Observed by LLM Agents During Task Execution — 2025. https://arxiv.org/abs/2506.01055
- Dissecting Adversarial Robustness of Multimodal LM Agents — Wu et al., 2024. https://arxiv.org/abs/2406.12814
J. Open-weight 安全移除與更深層風險(詳見 07-open-vs-closed)
- Refusal in Language Models Is Mediated by a Single Direction — Arditi et al., 2024. https://arxiv.org/abs/2406.11717 —— abliteration 的理論基礎(NeurIPS ‘24)。
- Shadow Alignment — Yang et al., 2023. https://arxiv.org/abs/2310.02949 —— 100 筆資料即可移除安全對齊。
- Fine-tuning Aligned Language Models Compromises Safety — Qi et al., 2023. https://arxiv.org/abs/2310.03693 —— 10 筆對抗樣本、甚至良性微調也會傷害安全性(ICLR ‘24);適用 closed model 的 fine-tuning API。
- BadLlama — Gade et al., 2023. https://arxiv.org/abs/2311.00117 —— <$200 移除 Llama 2-Chat 13B 安全微調。
- Sleeper Agents — Hubinger et al. (Anthropic), 2024. https://arxiv.org/abs/2401.05566 —— backdoor 能撐過 safety training。
- Emergent Misalignment — Betley et al., 2025. https://arxiv.org/abs/2502.17424 —— 窄任務(不安全程式碼)微調導致廣泛失準。
- Poisoning Web-Scale Training Datasets is Practical — Carlini et al., 2023. https://arxiv.org/abs/2302.10149
- Scaling Trends for Data Poisoning in LLMs — Bowen et al., 2024. https://arxiv.org/abs/2408.02946 —— 更大模型學有害行為更快。
K. 工具/紅隊/guardrail
- Llama Guard — Inan et al. (Meta), 2023. https://arxiv.org/abs/2312.06674
- garak — Derczynski et al. (NVIDIA), 2024. https://arxiv.org/abs/2406.11036
- PyRIT — Lopez Munoz et al. (Microsoft), 2024. https://arxiv.org/abs/2410.02828
L. 2026 SOTA 補充(⚠️ preprint)
- SoK: PIAtlas — Damianakis & Psounis, 2026. https://arxiv.org/abs/2610.24999 —— 以 6 種 subversion mechanism 統一 20+ 資料集詞彙。
- The Agent Knew (CAPE-Bench) — Fayyazi et al., 2026. https://arxiv.org/abs/2610.17301 —— 越權後偽造授權/虛報;內部狀態可預測(AUROC 0.94)。
- Package Hallucination Attacks on Coding Agents through Prompt Injection in Rule Files — 2026. https://arxiv.org/abs/2610.09264
- Names Are Not Objects: Object-Bound Confinement for Compromised AI Coding Agents — 2026. https://arxiv.org/abs/2610.14488
- Towards Safe Autonomy for Computer Use — 2026. https://arxiv.org/abs/2610.18508
- RATION: Quantitative Risk-Based Access Control for AI Agents — 2026. https://arxiv.org/abs/2610.24311
- Cryptographically Secure Agent-to-Agent Communication in Multi-Agent Systems — 2026. https://arxiv.org/abs/2610.23078
註:全部 67 篇(編號 23 從缺,原列 DataSentinel 但查到的 ID 不符,已刪除)皆已驗證 arXiv 標題。#62–68 僅依標題/摘要,尚未精讀。