The Set-and-Forget Trap
You install an MCP server, it works, and you move on with your life. Three months later, you're still running that exact version. In those three months, the server's dependencies might have accumulated known vulnerabilities, the model you're using might have changed how it handles tool calls, and the server's author might have fixed bugs and added features you're missing.
This isn't hypothetical. The JavaScript and Python package ecosystems discover new vulnerabilities daily. A dependency that was clean when you installed it might now have a critical CVE. If that dependency handles data that your MCP server processes, you're exposed without knowing it.
Security Drift
The supply chain risk of outdated dependencies is the most concrete cost. Every week you don't update, the probability that one of your dependencies has a known vulnerability increases. Most of these vulnerabilities are minor, but it only takes one critical one in a data-handling package to create a real problem.
Running npm audit or pip-audit on your MCP server's directory periodically takes less than a minute and tells you exactly where you stand. If the audit comes back clean, great. If it flags issues, you know what to update.
Compatibility Drift
AI clients (Claude Desktop, Cursor, etc.) update regularly. Sometimes these updates change how they interact with MCP servers. An older server version might not support a new protocol feature, might handle connection lifecycle differently, or might send responses in a format that the updated client handles less gracefully.
Compatibility issues usually show up as subtle problems rather than hard failures. The server still works, but some queries time out more often, or error messages are less informative, or connection drops happen more frequently. These degraded behaviors can be hard to diagnose because nothing is obviously broken.
A Practical Update Schedule
You don't need to update daily, but you should have a cadence. Monthly updates for actively-used servers is a reasonable baseline. Check for security advisories between scheduled updates. And when you hear about a significant MCP protocol change, prioritize updating your servers.
Before updating, check the changelog. Breaking changes in major version bumps might require configuration adjustments. Minor and patch updates are usually safe to apply directly. Keeping version-pinned lockfiles (package-lock.json, poetry.lock) means you can always roll back if an update causes problems.
Platforms that track security scores can alert you when a tool's grade drops, which often indicates new dependency vulnerabilities. This is more efficient than manually checking each server's dependencies.
Related Reading
- Supply Chain Risks in the AI Tool Ecosystem
- Why Dependency Transparency Matters for AI Tools
- Monitoring MCP Server Health in Production
Browse MCP servers on Skillful.sh. Search security-scored AI tools.