How MCP Servers Can Expose Enterprise Secrets

Recent analysis from Keeper Security highlights significant security gaps in the adoption of the Model Context Protocol (MCP), originally introduced by Anthropic. The study identifies several mechanisms through which MCP servers can inadvertently expose enterprise credentials, including storing plaintext tokens in configuration files and failing to enforce the principle of least privilege during deployment.

A critical concern involves the "mcp-remote" OAuth proxy, where vulnerabilities such as CVE-2025-6514 allow for operating system command injection and remote code execution. This specific flaw could enable attackers to steal credentials directly from the client machine running the proxy. To mitigate these risks, the authors recommend centralizing secret management, using short-lived rotating credentials, and maintaining strict visibility over all MCP servers operating within an enterprise environment.