MCP server security:what to check before launch
Updated · 4 min read
An MCP server gives an AI assistant access to your systems. That's convenient, but the model also reads content that neither you nor the user wrote: emails, tickets, web pages. Here are the real risks, as described in the MCP specification and by OWASP, and the checklist we go through before every launch.
The biggest risk: prompt injection
A model can't reliably tell the user's instructions from text that came back in a tool result. If someone hides an instruction in a ticket, an email or a product description, the model may follow it. OWASP puts prompt injection first on its list of LLM risks and admits there may be no fool-proof protection.
In May 2025 researchers at Invariant Labs showed this on GitHub's MCP server: a malicious issue in a public repository got the assistant to move data from a private repository into a public pull request. The server's code was fine. All it took was the user ticking "always allow."
The lethal trifecta
Simon Willison's name for three things in one conversation: access to private data, exposure to untrusted content and a way to send something out. With all three present, stealing data takes one successful attack. When we design a server, we check whether its tools add up to that combination, and if they do, we split them or require approval.
Poisoned and changing tools
A tool's description goes straight to the model. In April 2025 Invariant Labs showed an "add two numbers" tool whose description told the model to read SSH keys and the configuration of other servers. A server can also change its descriptions after the user has approved it. The MCP specification says outright that tool annotations from untrusted servers are only hints, and Anthropic notes that a Verified label in its directory is not a security audit.
Sign-in and tokens
- Every user signs in with their own account through OAuth. One shared API key for everyone means access with no record of who did what. Details in OAuth for MCP servers.
- The server accepts only tokens issued for it and doesn't pass them on. The MCP specification forbids it, because it bypasses rate limits, breaks the audit trail and turns the server into a proxy for stolen tokens.
- Scopes from the minimum up. No scopes like * or full-access, and wider permissions only once a tool needs them.
- Your system checks permissions on every call, not the model. A scope in the token isn't enough on its own.
- State handles (a cart ID, say) are not proof of identity. Bind them to the user from the token.
Tools and user approval
- Separate tools for reading and for changes. Claude's directory rejects a tool that does both.
- Every tool has a readOnlyHint or destructiveHint annotation. In ChatGPT it has to be an explicit value with a justification, and being able to undo a change doesn't make a tool non-destructive.
- Irreversible operations (payment, deletion, sending) always need a human to confirm. Claude doesn't accept connectors that transfer money into its directory.
- A tool returns only the data the task needs. Less data in the conversation means less to steal.
- No hidden instructions in tool descriptions. Claude's review rejects descriptions that steer the model or interfere with other tools.
Infrastructure
- Input validation on every tool and rate limits on calls. The MCP specification requires both.
- Logs of tool calls with the user's ID, without personal data in the content.
- Traffic restricted to ChatGPT's and Claude's IP ranges if the server isn't meant to be public.
- Up-to-date dependencies. In 2025 critical flaws turned up in mcp-remote (CVE-2025-6514, command injection through the authorization URL) and in the MCP Inspector (CVE-2025-49596, remote code execution), among others.
- Local servers: stdio, or a token over HTTP, never an open port without sign-in.
Checklist before launch
- Does any tool combine private data, untrusted content and sending data out?
- Does every tool that changes data need approval and carry an annotation?
- Does the server run with the signed-in user's permissions rather than an admin account?
- Does the backend check permissions, rather than the tool description?
- Do you know who called which tool and when?
- When tool descriptions change, does someone review them before deployment?
- Can prompt injection be blocked completely?
- No method works every time. That's why you limit the damage: the smallest possible permissions, human approval for important operations, and permissions checked in your system rather than by the model.
- Is a public server without sign-in safe?
- It can be, if it only exposes public read-only data such as a product catalog and prices. That's how search works in [Drukalo](/work/drukalo). Anything that touches a customer's account or changes data needs sign-in.
- Does a Verified label in Claude's directory guarantee security?
- No. Anthropic says it isn't a security audit, and the developer can change tools after review. Check the permissions a connector asks for and changes to its tools.
- MCP integration: your ERP, store and CRM in ChatGPT and Claude
- Aesthetic clinic websites in Poland
- Online booking for a private clinic
- An AI assistant for your clinic: what's allowed in 2026
- ChatGPT product feed: get your products into ChatGPT shopping
- Clinic advertising in Poland: what a website may say
- Your clinic in ChatGPT and Google
- Patients from abroad: a medical tourism website
- Claude connectors: what they are and how to add one
- MCP servers in ChatGPT: how to connect one
- How to connect WooCommerce to ChatGPT in 2026
- OAuth for MCP servers: sign-in in ChatGPT and Claude
- MCP Apps: interface in ChatGPT and Claude
- How to test an MCP server in ChatGPT and Claude
- MCP server or API? When you need both
- Hiring a software studio in Poland
- Your Shopify or WooCommerce store in ChatGPT
- What is an MCP server? Explained for online stores
- Claude connector for SaaS: what it takes
- How to sell in ChatGPT in 2026
- MCP server, ChatGPT plugin or Claude connector?
- How to get listed in ChatGPT and Claude
Want to connect your systems safely?
MCP servers