We've been scanning MCP (Model Context Protocol) servers to understand the security posture of the ecosystem. After 306 servers and 1,211 vulnerabilities, here's what we found:

- 69 critical vulnerabilities - 32 servers using eval() on untrusted input - 31 SQL injection vulnerabilities - 32 servers with hardcoded API credentials

10.5% of servers have at least one critical vulnerability.

MCP servers run with your permissions. AI agents trust them implicitly. A prompt injection can exploit them remotely.

MCPSafe (https://mcpsafe.org) lets you scan before you connect. Free tier includes 30 top servers.

Feedback welcome: What patterns are we missing? CI/CD integration useful?