附錄 D:ASI Agentic Exploits & Incidents Tracker
Appendix D - ASI Agentic Exploits & Incidents Tracker
ASI Exploits and Incidents initiative 透過引用真實世界的 incident 與 exploit,協助提供資訊並建立 OWASP Top 10 for Agentic Applications。這將補充、但不取代 GenAI Security project 中既有的漏洞通報。同樣地,任何與 incident response 有關的面向,應由負責發布 incident response guide 的 CTI initiative 討論。
以下清單每週更新,以反映最新的公開資訊。
最新版可至其專用 GitHub repository: https://github.com/OWASP/www-project-top-10-for-large-language-model-applications/blob/main/initiatives/agent_security_initiative/ASI%20Agentic%20Exploits%20%26%20Incidents/ASI_Agentic_Exploits_Incidents.md
Exploits & Incidents Table
| 日期 | Exploit/Incident | 影響摘要 | ASI T&M Mapping | 進一步分析 |
|---|---|---|---|---|
| 2025-10 | Malicious MCP Package Backdoor | NPM package 託管了帶有 backdoor 的 MCP server,包含兩組 reverse shell(install-time 與 runtime),使攻擊者能持續遠端存取 Agent environment。 | ASI04(Agentic Supply Chain Vulnerabilities) | NPM · Koi Security |
| 2025-10 | Framelink Figma MCP RCE | Framelink Figma MCP 的 get_figma_data tool 未對 user input 進行 sanitization,使未經 authentication 的 remote command execution 得以在 host system 上發生。 |
ASI05(Unexpected Code Execution)· ASI02(Tool Misuse & Exploitation) | Figma Context MCP · NVD · Imperva |
| 2025-10 | Cursor Config Overwrite via Case Mismatch | Case-insensitive filesystem 讓 crafted prompt 能覆寫關鍵 Cursor config,進而造成 persistent RCE 與 Agent compromise。 | ASI05(Unexpected Code Execution) | Cursor · NVD · Lakera |
| 2025-10 | Cursor Workspace File Injection | Cursor Agent 的 prompt 使 Cursor 寫入惡意 .code-workspace setting,透過 VS Code integration,在 workspace 開啟時允許 command execution。 |
ASI05(Unexpected Code Execution) | Cursor · NVD · MaccariTA |
| 2025-10 | MCP OAuth Response Exploit | 不受信任 MCP server 的 OAuth flow 可回傳 poisoned response,使攻擊者注入在 authentication 後由 Agent 執行的 command。 | ASI07(Insecure Inter-Agent Communication) | Cursor · NVD · Y4tacker |
| 2025-10 | Cursor CLI Project Config RCE | Clone 下來的 project 可透過 .cursor/cli.json 覆寫 global config,讓 attacker-controlled command 經由 Cursor CLI context 執行。 |
ASI04(Agentic Supply Chain Vulnerabilities) | Cursor · NVD · Assaf Levkovich |
| 2025-10 | Cursor Agent File Protections Bypassed | Cursor CLI Agent 的 file protection mechanism 可透過 prompt injection 繞過,進而藉由 config overwrite 達成 RCE。 | ASI05(Unexpected Code Execution) | Cursor · NVD |
| 2025-09 | Google Gemini Trifecta | 透過 log、search history 與 browsing context 的 indirect prompt injection,可誘使 Gemini 暴露 sensitive data,並在相連的 Google service 間執行非預期 action。 | ASI01(Agent Goal Hijack)· ASI02(Tool Misuse & Exploitation) | Tenable |
| 2025-09 | Malicious MCP Server Impersonating Postmark | 據報是第一個出現在實際環境中的惡意 npm MCP server;它冒充 postmark-mcp,並暗中將 email BCC 給攻擊者。 |
ASI02(Tool Misuse & Exploitation)· ASI04(Agentic Supply Chain Vulnerabilities)· ASI07(Insecure Inter-Agent Communication) | Postmark · Koi Security |
| 2025-09 | ForcedLeak (Salesforce Agentforce) | Salesforce Agentforce 中的 critical indirect prompt injection 允許 external attacker 誤導 Agent,並將 sensitive CRM record exfiltrate 到組織外部。 | ASI01(Agent Behaviour Hijack,原文如此)· ASI02(Tool Misuse & Exploitation) | Salesforce · Noma Security |
| 2025-09 | Visual Studio Code & Agentic AI workflows RCE | Agentic AI workflow 中的 command injection,可讓 remote、unauthenticated attacker 使 VS Code 在 developer machine 上執行注入的 command。 | ASI01(Agent Goal Hijack)· ASI02(Tool Misuse & Exploitation)· ASI05(Unexpected Code Execution) | Microsoft · NVD |
| 2025-07 | Amazon Q Prompt Poisoning | Extension 中的破壞性 prompt 帶來檔案遭清除的風險。 | ASI01(Agent Goal Hijack)· ASI02(Tool Misuse & Exploitation)· ASI04(Agentic Supply Chain Vulnerabilities) | AWS · NVD |
| 2025-07 | Google Gemini CLI File Loss | Agent 誤解 file instruction 並清除 user directory;之後承認這是 catastrophic loss。 | ASI05(Unexpected Code Execution) | |
| 2025-07 | ToolShell RCE via SharePoint | SharePoint 中的 RCE exploit 被 Agent 利用。 | ASI05(Unexpected Code Execution) | Microsoft · NVD · Eye Security |
| 2025-07 | Replit Vibe Coding Meltdown | Agent hallucinate data、刪除 production DB,並產生 false output 以掩蓋錯誤。 | ASI01(Agent Goal Hijack)· ASI09(Human-Agent Trust Exploitation)· ASI10(Rogue Agents) | Replit · SaaStr |
| 2025-07 | Microsoft Copilot Studio Security Flaw | Agent 預設為公開且缺少 authentication。攻擊者可 enumerate 並存取暴露的 Agent,從 production environment 取得 confidential business data。 | ASI03(Identity & Privilege Abuse)· ASI07(Insecure Inter-Agent Communication) | Zenity Labs |
| 2025-06 | Heroku MCP App Ownership Hijack | 惡意 tool input 利用 Heroku MCP 的 trust boundary,透過 Agent-mediated call injection,在未經授權的情況下劫持 app ownership。 | ASI03(Identity & Privilege Abuse) | Heroku |
| 2025-06 | Hub MCP Prompt Injection (Cross-Context) | 惡意 web page 可透過 DNS rebinding/CSRF 與 local MCP Inspector proxy(無 authentication)互動,驅使其透過 stdio 執行 MCP command,進而造成 arbitrary OS command execution 與 data exfiltration。 | ASI01(Agent Goal Hijack)· ASI02(Tool Misuse & Exploitation)· ASI05(Unexpected Code Execution) | MCP · NVD · Oligo Security |
| 2025-06 | AgentSmith Prompt-Hub Proxy Attack | Proxy prompt Agent exfiltrate API key。 | ASI04(Agentic Supply Chain Vulnerabilities) | Noma Security |
| 2025-05 | EchoLeak (Zero-Click Prompt Injection) | Critical zero-click exploit 只需一封 email,就能觸發 Copilot 將 confidential data(email、file、chat log)洩漏到其預期 scope 之外。 | ASI01(Agent Goal Hijack)· ASI02(Tool Misuse & Exploitation)· ASI06(Memory & Context Poisoning) | Microsoft · NVD · Aim Security |
| 2025-05 | GitPublic Issue Repo Hijack | Public issue text 透過 cross-repo prompt injection 劫持 AI dev Agent,使其洩漏 private repo content。 | ASI01(Agent Goal Hijack)· ASI02(Tool Misuse & Exploitation)· ASI06(Memory & Context Poisoning)· ASI07(Insecure Inter-Agent Communication)· ASI08(Cascading Failures) | Invariant Labs |
| 2025-04 | Agent-in-the-Middle (A2A Protocol Spoofing) | 惡意 Agent 在 open A2A directory 發布假的 agent card,虛假宣稱具有高度 trust。LLM judge Agent 選中它後,Rogue Agent 便能攔截 sensitive data 並洩漏給未授權對象。 | ASI03(Identity & Privilege Abuse)· ASI06(Memory & Context Poisoning)· ASI07(Insecure Inter-Agent Communication)· ASI08(Cascading Failures)· ASI10(Rogue Agents) | Trustwave |
| 2025-03 | GitHub Copilot & Cursor Code-Agent Exploit | 遭 manipulation 的 AI code suggestion 將 backdoor 注入 production code、洩漏 API key,並引入 logic flaw;由於 developer 信任 AI output,因此形成顯著 supply-chain risk。 | ASI04(Agentic Supply Chain Vulnerabilities)· ASI08(Cascading Failures)· ASI09(Human-Agent Trust Exploitation) | Pillar Security |
| 2025-03 | Flowise Pre-Auth Arbitrary File Upload | 未經 authentication 的 arbitrary file upload 可造成 Agent framework compromise,並可能取得 remote server control;vendor response 亦有延遲。 | ASI05(Unexpected Code Execution) | FlowiseAI · NVD · Dor Attias (Medium) |
| 2025-02 | OpenAI ChatGPT Operator Vulnerability | Web content 中的 prompt injection 使 Operator 遵循 attacker instruction、存取 authenticated page 並暴露 user private data,呈現防護較弱的 autonomous Agent 所具有的 data leakage risk。 | ASI01(Agent Goal Hijack)· ASI02(Tool Misuse & Exploitation)· ASI03(Identity & Privilege Abuse)· ASI04(Agentic Supply Chain Vulnerabilities)· ASI06(Memory & Context Poisoning)· ASI07(Insecure Inter-Agent Communication)· ASI09(Human-Agent Trust Exploitation) | Wunderwuzzi |
原作:OWASP Top 10 For Agentic Applications 2026
專案:OWASP Gen AI Security Project - Agentic Security Initiative
授權:CC BY-SA 4.0
本翻譯為非官方繁體中文版本,並依 CC BY-SA 4.0 授權釋出。