By Adrian Goins, Obot AI
757 MCP servers compromised. 36% scored failing grades. Zero earned an A.
Those aren't projections — that's what a recent audit of real-world MCP OAuth implementations found. And the culprit isn't sophisticated attacks. It's shortcuts: client secrets hardcoded in frontend code, redirect URIs left wide open, PKCE skipped because it seemed like overkill.
OAuth is well understood at this point. So why are MCP implementations getting it so consistently wrong? Adrian Goins breaks down the specific patterns that are turning OAuth from a security control into a liability — and what a secure MCP OAuth implementation actually looks like.
Originally published on Obot AI.
Top comments (0)