Are MCP Servers Safe? Risks and a Simple Checklist
Updated November 5, 2026
Short answer: MCP servers can be safe, but none is automatically safe. A local server runs code on your computer, a remote server sees what you send it, and text returned by any tool can try to mislead your assistant. Use servers from people you trust, give them the least access that works and keep approval on for anything that changes data.
What are the main risks?
Local servers run code on your computer
A local server runs with your user's permissions. The official MCP security guidance warns that a malicious package or start command could read files such as SSH keys or send data elsewhere. Treat a local server like any other software you install.
Remote servers see what you send
When your assistant calls a hosted server, that provider receives the request. Don't send private data to a service you wouldn't otherwise trust with it.
Tool results can carry instructions
This is called prompt injection. A web page, email or tool result can include hidden text that tries to steer your assistant. The risk is highest when one assistant can read untrusted content and also take actions such as sending messages.
Too much access
A broad token gives an assistant, and anyone who steals that token, more reach than it needs. The MCP guidance recommends starting with minimal permissions and adding more only when required.
A checklist before you connect a server
- Who runs it? Prefer official servers from the company whose product it connects to.
- Can you read the source? Open source lets you, or someone else, check what it does.
- Read the tool list. Does it include tools that delete, send or spend?
- Read the start command. Know exactly what you are running, and pin a version where you can.
- Limit access. Use read-only credentials and restricted API keys.
- Keep approval on. Don't auto-approve tools that change anything.
- Try it on test data first. Never start with production.
- Remove what you stop using. Fewer connections means fewer risks.
What does "verified" mean on MCP Collection?
It means a hosted server answered a real MCP handshake in our daily check. It does not mean we audited its code or that it is safe to use. We reject submissions that don't work or aren't MCP servers, but you should still check each one yourself.
Our security server picks include tools that scan for prompt injection and look up a server's history.
Frequently asked questions
Can an MCP server steal my data?
A malicious one could. A local server can reach what your user account can reach, and a remote one receives what you send it. That is why the source and the operator matter.
What is tool poisoning?
Tool poisoning is when a server hides instructions in its tool descriptions to influence the assistant. Your app sends those descriptions to the assistant, so a bad server can try to steer it before any tool is called.
Does a verified badge mean a server is safe?
No. On this site it means the server responded correctly to a live check. It says nothing about the code or the operator.