MCP Security
An MCP server can read your data and act on your behalf, so a bad one, or a good one that's been tricked, can do real damage. These guides cover the attacks that actually happen and the defences that work, for people using servers and for people building them.
Start with MCP security best practices.
Security best practices
The main security risks of MCP servers for people who connect them and people who build them, with a checklist for each and links to the spec.
Tool poisoning and rug pulls
What tool poisoning and rug pulls are in MCP, how security researchers demonstrated them, and the checks that catch them before a server reaches your model.
Prompt injection
How hidden instructions in tool results can steer an AI agent into leaking data, with a tested example and defences for users and builders.
Common vulnerabilities
The attacks the MCP spec names, including confused deputy, token passthrough and SSRF, with the fix for each and the real CVEs.
Sandbox local servers
Run a local MCP server in a locked-down Docker container with no network, a read-only file system and one writable folder, then connect it to Claude Desktop.
Evaluate a server
A checklist for judging an MCP server before you install it, covering the publisher, the pinned version, its tool list, its access, and what scanners can tell you.