Malicious MCP server
Also known as: bad MCP ยท MCP attack
A malicious MCP server is a Model Context Protocol server that presents legitimate-looking tools but behaves maliciously โ exfiltrating, misdirecting, or poisoning.
- Looks like a normal integration.
- Exposes tools that leak data or mislead the model.
- A supply-chain vector specific to the agent ecosystem.
The check: before wiring in an MCP server, audit what tools it exposes and what data they can reach. An MCP server is code you're giving agent-level trust to.