工具增强型大语言模型(LLM)代理创建了新的供应链漏洞:MCP工具像第三方包一样安装,但其输出可能进入代理的推理上下文。这导致了confused-deputy风险,即攻击者控制的输入导致原本无害的工具行使合法权限,如文件、环境变量或网络操作,并将敏感或指令类内容反映到LLM可见的字段中。我们提出了SandScope,一个MCP-aware审计框架,结合运行时见证检测与语义工具配置文件。SandScope在WebAssembly System Interface(WASI)下执行可移植工具或驱动未修改的MCP服务器通过标准输入输出(stdio),提取LLM可见的sink,并报告可审计的源到sink见证,同时分别记录网络意图和出站证据。其语义层从工具/列表元数据和静态注册中恢复声明的能力,以表征执行不完整时的攻击面。我们评估了SandScope在受控的跨语言项目、规避基准和100个仓库MCP语料库中的表现。SandScope完成了35个仓库的浅层动态扫描,并通过更广泛的语义配置文件扫描,恢复了71个仓库中1,127个工具的元数据,包括886个具有安全敏感声明能力的工具。通过针对35个动态扫描仓库的模式引导探索扫描,重新执行了33个并观察到12个源到sink见证。这些结果表明,SandScope通过受控执行、MCP-aware sink提取、运行时见证报告和语义能力配置文件,为MCP工具风险提供了实用且可审计的证据。
英文摘要:
Tool-augmented Large Language Model (LLM) agents create a new supply-chain surface: Model Context Protocol (MCP) tools are installed like third-party packages, yet their outputs can enter the agent's reasoning context. This enables confused-deputy risks in which attacker-controlled inputs cause otherwise benign tools to exercise legitimate authority over files, environment variables, or network-facing operations and reflect sensitive or instruction-like content into LLM-visible fields. We present SandScope, an MCP-aware audit framework that combines runtime witness detection with semantic tool profiling. SandScope executes portable tools under WebAssembly System Interface (WASI) or drives unmodified MCP servers over standard input/output (stdio), extracts LLM-visible sinks from tool results and prompt/message fields, and reports auditable source-to-sink witnesses from environment, file, and tool-input sources while separately recording network-intent and egress evidence. Its semantic layer recovers declared capabilities from tools/list metadata and static registrations to characterize attack surface when execution is incomplete. We evaluate SandScope on controlled cross-language subjects, an evasion benchmark, and a 100-repository MCP corpus. SandScope completes shallow dynamic scans for 35 repositories and, through a broader semantic profiling pass, recovers metadata for 1,127 tools across 71 repositories, including 886 tools with security-sensitive declared capabilities. A schema-guided exploration pass over the 35 dynamically scanned repositories re-executes 33 and observes source-to-sink witnesses in 12. These results show that SandScope provides practical, auditable evidence for MCP tool risk through controlled execution, MCP-aware sink extraction, runtime witness reporting, and semantic capability profiling.
Runtime Authorization Consistency Checking for MCP-based Agentic Workflows
Institute of Computing Technology, Chinese Academy of Sciences(中国科学院计算技术研究所);University of Chinese Academy of Sciences(中国科学院大学);Fuxi Institution(伏羲研究院);Center for Internet Governance, Tsinghua University(清华大学互联网治理研究中心)
ETDI: Mitigating Tool Squatting and Rug Pull Attacks in Model Context Protocol (MCP) by using OAuth-Enhanced Tool Definitions and Policy-Based Access Control
Amazon Independent Researcher(亚马逊独立研究者);Project Kuiper Security (KPES)(项目Kuiper安全);Proactive Security OWASP, Amazon Web Services(主动安全OWASP,亚马逊网络服务);Intuit Adversarial AI Security reSearch (A2RS)(Intuit对抗AI安全reSearch)
本文提出 ETDI 安全扩展,通过密码学身份验证、不可变版本化工具定义、OAuth 2.0 权限管理及基于策略的细粒度访问控制,缓解 MCP 中的工具投毒和 Rug Pull 攻击,构建更安全可控的 AI 工具交互生态。
A2M: Trace-Optimized Agent Hijacking in the MCP Ecosystem
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences(中国科学院深圳先进技术研究院);University of Chinese Academy of Sciences(中国科学院大学);Nanyang Technological University(南洋理工大学);Southern University of Science and Technology(南方科技大学);Shenzhen University of Advanced Technology(深圳先进技术学院);University of Macau(澳门大学)
Systematization of Knowledge: Security and Safety in the Model Context Protocol Ecosystem
Tribhuvan University(特里布文大学);University of North Dakota(北达科他大学);Youngstown State University(亚当斯州立大学);University of Missouri(密苏里大学);University of Toledo(托莱多大学)
An Organization-Scoped LLM Agent Runtime Architecture for Regulated Cybersecurity Operations
Innov-Acts Ltd(Innov-Acts有限公司);Dept. of Digital Systems University of Piraeus(数字系统系希腊比雷埃克斯大学);Harokopio University(哈罗基奥大学);Dept. of Informatics and Telematics(信息学与电信系)
Persistent Billable State: Denial-of-Wallet Attacks and Defenses in Tool-Calling LLM Agents
Institute of Information Engineering, Chinese Academy of Sciences(中国科学院信息工程研究所);School of Cyber Security, University of Chinese Academy of Sciences(中国科学院大学网络空间安全学院);Beihang University(北京航空航天大学);Beijing University of Posts and Telecommunications(北京邮电大学)
ChainCaps: Composition-Safe Tool-Using Agents via Monotonic Capability Attenuation
Independent Researcher, Seattle, WA, USA(华盛顿州塞勒姆独立研究员);Independent Researcher, New York City, NY, USA(纽约市纽约独立研究员);King Abdullah University of Science and Technology(国王阿卜杜勒阿齐兹科学技术大学)
APEX: Active Protection at Execution Boundaries for LLM Agents
Tsinghua University(清华大学);Imperial College London(帝国理工学院);Nanjing University(南京大学);The Chinese University of Hong Kong(香港中文大学);University College London(伦敦大学学院);A*STAR(新加坡科技研究局);University of British Columbia(不列颠哥伦比亚大学);Shenzhen University(深圳大学);The University of Hong Kong(香港大学)