The official MCP Python SDK let a malicious server collect your AI agent's OAuth secret, and upgrading is only half the fix
On September 28 the maintainers of the official Model Context Protocol Python SDK published GHSA-qx49-fqc8-xw99: for sixteen months, an MCP server an AI agent connected to could decide where the agent's OAuth credentials were sent, including the client secret it uses to log in to a real service.