MCP Security Risks

Every connected MCP server is a piece of untrusted text in context.

Four types of risks:

tool description poisoning – malicious instructions in description: injection triggering every session;
malicious or compromised servers – supply chain attack;
tool shadowing – interception of call with sensitive parameters;
credential risks – OAuth tokens, API keys.

Mitigation: audit descriptions as untrusted input, server version pinning, least privilege, independent verification with Sidecar.