MCP 安全风险与利用链:不可信服务器与困惑的客户端
安全研究者 Johann Rehberger 详细分析 MCP 的安全风险,指出仅将恶意工具服务器加入对话上下文即可劫持会话,无需实际调用工具。他演示了在工具元数据中嵌入隐藏 Unicode Tags 指令的端到端利用链,使 Claude 在用户界面看不到异常的情况下先调用 enter_matrix 工具,再执行用户请求的 message_of_the_day。作者称该 Claude 遵循隐藏 Unicode Tag 指令的问题早在 14 个月前就已负责任披露给 Anthropic,但未被视为安全漏洞,此次 MCP 弱点也在数周前告知 Anthropic 且仅得到可复现的回复。他建议不要随意连接不可信 MCP 或 OpenAPI 工具服务器,并检查代码、接口定义与隐藏指令。
they all suffer from prompt injection and confused deputy threats. Many of the exploits we have discussed in the past involved tool invocation, from Zapier and ChatGPT exploits to Microsoft Copirate exploit to search email before leaking them, all of these required to know what tools are accessible and how to invoke them. So I figured, I spend some time explaining how I go about debugging and discovering things, in this case with MCP. Anthropic discusses prompt injection and other risks in their
匹配字段正文