A report from Hush Security, titled The State of MCP Configuration: The Identity Security Gaps, has identified a widespread practice of embedding credentials directly into Model Context Protocol (MCP) configuration files. These files, which are used to configure AI coding assistants, are often committed to public GitHub repositories, leaving API keys and access tokens exposed to anyone who stumbles upon them.
The findings highlight a common but dangerous shortcut: developers hardcoding secrets into configuration files for convenience, then pushing those files to public repos without proper sanitization. Because MCP configurations are designed to be shared and reused, the exposure is not limited to a single project—anyone who clones or forks the repository can extract the credentials.
Hush Security's research underscores the identity security gap in the AI tooling ecosystem. Unlike traditional secrets management, MCP configurations are a relatively new attack surface, and many developers may not yet treat them with the same caution as other code. The report suggests that organizations adopting AI coding tools should audit their public repositories for hardcoded credentials and implement automated scanning to prevent future leaks.