Is MCP safe? Risks and how to limit them
Is MCP safe? MCP, short for Model Context Protocol, is a standard way to connect an AI app such as Claude or ChatGPT to your business software, so that the AI can look things up and make changes there. MCP itself only passes requests back and forth. So MCP security depends on what you connect and how much access you give. With narrow, personal access, the risk is manageable.
MCP is a door, and you hold the key
Think of giving someone a key to your office. A key for the front door and one room is a small risk. A master key for every room and the safe is a large one. The door is the same in both cases.
MCP works like that. The AI app connects with an access key, a kind of password for programs. The AI can do exactly what that key allows. With an administrator's key it can do everything an administrator can. With a key for reading in one area, that is all it can do. What is new is who uses the key: an AI that decides for itself which step to take next.
Four risks and how to limit each
| Risk | What it means | How to limit it |
|---|---|---|
| Too much access | The AI can change or delete far more than the task needs | Reading only at first. Allow changes area by area |
| Unknown connectors | A connector from an unknown author gets your rights | Use connectors from the maker of your software |
| Hidden instructions | A file or email contains a line aimed at the AI | Let the AI app ask you before it changes anything |
| Lost keys | A key is pasted into a chat or left in a shared file | Store keys safely. Cancel a key at once when in doubt |
A connector is the small program that links the AI app to one piece of software. In MCP it is called a server. The trick with hidden instructions is known as prompt injection and has its own article.
Two more things to check
- Where your data goes. Whatever the AI reads is processed by the company behind the AI app. Check its terms on data, and use a business plan that fits your rules.
- Whether the AI app asks first. AI apps differ here. Where you have the choice, keep the question "May I change this?" switched on, at least in the first weeks.
A careful order for the first month
- One person, one AI app, reading only. Ask questions and compare the answers with what you see in your software.
- Allow creating and changing in one area, such as tasks. Keep deleting switched off.
- Add more people, each with their own key. Never share one.
- Review the connections every few months. Remove the ones nobody uses.
So, is it safe enough?
Yes, for a company that chooses its connectors with care, keeps access personal and narrow, and can see what was done. The risk grows with every connector from an unknown source and every right given "just in case". Both are choices you can take back at any time.
How Worknano handles this
In Worknano, MCP stays off until an admin turns it on for the workspace. Each person connects with a personal key. Every request is checked against what the workspace allows over MCP, what that connection may touch and the person's own role. Admins decide per AI app what it can read, create or update. Deleting is off by default, keys expire after 90 days, and Disconnect cuts access at once. Secrets, users and roles are never exposed. Other passwords and keys belong in Secrets, the built-in vault that records who viewed or copied an entry.
- MCP in Worknano: your permissions, not a master key
- Security at Worknano
- Secrets: the password vault in your workspace
Create a free workspace and connect your AI app with reading rights only.