Direct Answer
The main MCP risk surface includes: exposed API keys, unauthorized party authentication, rate-limit bypass, and session data leakage. The official specification mitigating controls: server authentication, TLS???everywhere transport encryption, and per???tool permission scoping. Best practices: periodic tool???permission review, access???log monitoring, and rate???limit enforcement on public servers.
Key Security Risks
1. Exposed API Keys: Servers that explicitly define tools may leak external data???access keys. Solution: Use environment Secrets and rotate periodically. 2. Weak Authentication: Failure to verify client identity may enable man???in???the???middle attacks. Solution: Token???based authentication + TLS. 3. Rate-Limit Bypass: Sequential requests may overload servers or consume resources. Solution: X-Rate-Limit-Remaining headers + exponential back???off. 4. Session Data Leakage: Unprotected Cookie headers may be stolen. Solution: HTTP???Only + SameSite=Strict headers.
Best Practices
??? Review tool permissions every 30 days and deactivate unused tools ??? Encrypt all communication headers via TLS 1.2+ ??? Apply rate limiting on every public endpoint ??? Rotate Secrets quarterly ??? Monitor access logs weekly ??? Use two???factor authentication (2FA) for server???admin access ??? Isolate MCP tools in separate subdomains to reduce attack surface
Specification Compliance
The practices above comply with modelcontextprotocol.io spec v1.2, including server???auth, transport encryption, and permission scoping. Non???compliant servers receive a lower confidence score in our directory (see: confidence formula 15% verification depth).
Examples from Our Directory
??? [microsoft/playwright-mcp](/en/mcp/microsoft-playwright-mcp): score 95/100 ??? includes server authentication + enforced TLS. ??? [googleapis/genai-toolbox](/en/mcp/googleapis-genai-toolbox): score 95/100 ??? headers locked + automatic secret rotation. ??? [cocoindex-io/cocoindex-code](/en/mcp/cocoindex-io-cocoindex-code): score 85/100 ??? lacks secret-rotation documentation (in progress).
FAQ
1. Is MCP less secure than traditional function calling? No ??? with best practices, MCP has documented, auditable security controls. 2. Does every server require client authentication? Public servers do; internal/private servers may not. 3. How do I monitor MCP vulnerabilities in my server? Weekly access logs + rate???limit monitoring tools. 4. Does MCP store user data? MCP itself does not store data; the hosting application is responsible for secure storage.
Decision Summary
For developers starting with MCP: learn best practices from day one (TLS, Secrets, permission review). For current developers: run a security audit using the checklist above, then assess your servers via our decision engine for vulnerability status.