安全研究发现,ChatGPT在同时启用两个连接器的对话场景中,存在一种未经用户明确告知的静默行为:模型会将后续的工具调用切换到不同的连接器上执行 1。具体表现为,一次性确认代码由连接器A生成,却通过连接器B发送,这种跨连接器的操作导致OAuth客户端不匹配,进而使服务器拒绝该调用 1。
由于每个连接器被视为拥有独立OAuth授权的单独应用程序,这种静默的数据切换实际上构成了未经用户同意的第三方数据传递,揭示了跨应用数据传输的安全风险 1。研究指出,仅共享MCP URL不足以将两个连接器视为同一实体,系统边界应严格基于OAuth客户端进行划分 1。正确的预期行为应是工具调用始终通过启动流程时的同一连接器和OAuth客户端进行 1。针对此漏洞的赏金项目目前要求提交者提供身份验证信息以完成流程 1。
A technical report highlights that when two connectors are enabled in a ChatGPT conversation, the model silently switches subsequent tool calls to a different connector without notifying the user 1. This behavior resulted in a specific failure where a one-time confirmation code generated by Connector A was sent via Connector B, causing the server to reject the request due to an OAuth client mismatch 1. The incident demonstrates how data can be passed across applications without explicit user consent, a risk that is particularly severe when third-party connectors are involved 1.
The analysis argues that each connector should be treated as an independent application with its own OAuth authorization, meaning that sharing a Model Context Protocol (MCP) URL does not justify treating them as a single entity 1. Instead, boundaries must be defined by the OAuth client rather than the URL structure 1. Consequently, the expected behavior is for all tool calls to remain within the same connector and OAuth client that initiated the workflow 1.
评论
还没有评论,欢迎留下第一条。