MCP security covers the risks of connecting an AI agent to tools through the Model Context Protocol (MCP). The tool descriptions and results a server sends are untrusted text that the model may follow as instructions.
How MCP works
MCP is an open standard for connecting AI applications to tools and data. A server offers tools, each with a name, a description and an input schema. The application shows the descriptions to the model, sends its calls to the server and returns the results.
Risks
- Tool poisoning. A description is text the model reads and the user rarely sees. A malicious server can hide instructions in it, or change it after approval.
- Injection through results. A tool's result is untrusted text too: a web page or a ticket an attacker wrote can carry an indirect prompt injection.
- Broad permissions. A server often holds a token for a whole mailbox or every repository, and one injected step can use all of it. See excessive agency.
Controls
- Vet servers. Install only ones you trust, pin versions and review tool descriptions when they change.
- Least privilege. Give each server a narrow token, read-only where possible, and connect only the servers the task needs.
- Check calls and results. Check each call's arguments before it runs, and scan each result before the model reads it.
With White Circle
Control checks tool calls before they run and blocks injected instructions before they reach a tool.
