Securing the Integration Protocol

The MCP Companion Guide provides actionable cybersecurity guidance for the protocol layer that mediates AI model and agent execution. It highlights that authorization must be enforced outside the model, recommends securing transports and endpoints (TLS, origin validation, DNS rebinding protections), calls out governance against capability drift (allowlists, signed artifacts, baselines), and prescribes hardening of tools, resources, and prompts (schema validation, sanitization, logging identifiers). The guide positions these measures as essential to ensure AI-driven actions remain auditable, least-privileged, and resistant to malicious or unintended behavior.