An MCP server lets your AI assistant use something outside the chat: your files, your mail, your website, a person. That is the point, and it is also the risk. Once a server is connected, your assistant can act for you, so the question is not whether MCP is safe in general. It is whether this server, with these permissions, is safe for you. We run humans fix ai, and we publish an MCP connector, so we have a stake in the answer. Every claim below comes from Anthropic's, OpenAI's or the MCP project's own pages, which we read on 15 November 2026.
What can go wrong
- An untrusted server can take what you share. OpenAI's guide warns that an untrusted server may access or steal information shared while you use it, or trick ChatGPT into calling tools in ways you did not intend, including changing or deleting data.
- Hidden instructions. Anthropic's help page says a malicious server may hide instructions that try to make Claude do something unintended. Claude has defenses against this, but the page says they only attempt to block such attacks.
- A server can change after you add it. The same page notes that a developer can change a tool's behavior without warning.
- The assistant acts as you. A connected service can let Claude read, create, change or delete data and take actions on your behalf.
- Permissions that are too wide. The MCP project's own security guidance names over-broad permissions as a risk, and notes that long lists of requested permissions are easy to approve without reading.
The checklist
- Know who runs it. Connect only to servers built and hosted by organizations you trust. Anthropic's page says custom connectors are for services it has not verified.
- Sign in, never share a password. A proper sign-in goes through the service's own page, so your assistant never sees your password. Check the address of the page you are on.
- Read the permissions on the sign-in page. Allow what the task needs. Limit or decline the rest.
- Keep approvals on for anything that changes things. OpenAI's guide says actions that write data ask for your confirmation by default. Check what the assistant is about to send before you press confirm, because a wrong write can destroy, change or share data.
- Be careful with "allow always" and "remember". Anthropic says to choose "Allow always" only for a server and a tool you trust to run without you watching. OpenAI says a remembered approval lasts for that conversation only.
- Turn off tools you do not need. Fewer tools means fewer ways to go wrong. Anthropic also advises switching off tools that can write to other apps before you use Research.
- Look at what goes in and out. Both assistants let you read the details of each tool call. Do it when something looks odd.
- Treat a local server like installing software. If a setup asks you to run a command on your computer, the MCP project's guidance says the exact command should be shown to you first, with a warning for risky ones such as
sudo. Do not run what you do not understand. - Remove what you stop using. You can revoke access in your assistant's settings or in the service itself.
- Report a server that behaves badly. Anthropic asks for reports of malicious servers through its vulnerability disclosure program.
How our connector handles this
You sign in with Google on our page, and your assistant never sees your password. It can save a draft request, but nothing reaches a developer until it has shown you the exact text and you have said yes. It never handles payment, and it only sees your own requests, not anyone else's. You can disconnect it at any time in Settings, and we email you whenever a new assistant is connected, so you notice one you did not expect. That does not make any connection risk-free, so the checklist above applies to ours too. See how it works and how to connect it.
If you are not sure
If an assistant connected something and you cannot tell whether it did what you wanted, or you want someone to look over what it changed, ask a developer. The first call is free. For servers that help you check your AI's work, see our list of the best MCP servers for non-developers.
Sources
Getting help takes about a minute
No hiring, no back and forth. A real person looks at your app.
-
1
Describe it
Tell us what you need in plain words. You don't have to be technical or have an account to start.
-
2
Free 15-minute call
A verified developer calls you. Many tasks are sorted right there, and it costs nothing.
-
3
Pay after you approve
If there is more to do, you get one fixed price up front. You pay only once you are happy.
Real people, reviewed by hand
Developers who do this every day
Every developer is approved by our team before they can take a task. Describe yours and the first one who is free picks it up.
Get help