Model Context Protocol servers are quickly becoming the default way to connect AI agents to real systems and when it connects an agent it becomes a new trust boundary, and the security decisions made at that boundary determine whether the agent is a helpful tool or an open door.

This talk shares practical security lessons from building an MCP server used by more than one team in production, grounded in the guidance I helped shape as a contributor to the OWASP GenAI Security Project. It is aimed at builders who are writing or planning to write their own MCP servers and want to get the security fundamentals right from the start, without needing a security background to follow along.

We will walk through the security considerations that matter most when building an MCP server. Attendees will leave with a clear mental model of the MCP attack surface, a short checklist of security practices to apply before going to production, and the confidence to build MCP servers that other people can trust. The goal is to help the agentic AI ecosystem mature by making secure-by-default MCP development something every builder understands, not just security specialists.