When your team builds its own connectors, platform guidance covers authentication, tool annotations such as read-only hints, input validation, and confirmation for irreversible actions. Treat every argument the model supplies as untrusted input, because a prompt injection can choose those arguments.
Read in this order
- OfficialSecurity and Privacy (OpenAI Plugins)Start here: the platform's security and privacy rules for connector builders.
- OfficialAuthentication (OpenAI Plugins)How to wire authentication so a connector acts only for the right user.
- OfficialBuilding MCP servers for plugins and API integrations (OpenAI API)How to build MCP servers that plug into the API and plugins.