If you ship or consume Model Context Protocol (MCP) servers, there are three findings worth checking against your own setup. I've been running a passive security scanner against the public MCP cohort for four weeks. Here's what the data says.
PKCE S256 not advertised (AUTH-003)
The OAuth 2.1 metadata document at /.well-known/oauth-authorization-server should include "code_challenge_methods_supported": ["S256"].
90%+ of public MCP servers advertise plain instead. The MCP spec (2026-07-28) treats plain as no PKCE support, and spec-compliant clients like Claude Desktop refuse to proceed with the flow.
Fix: one line in your metadata document.
No Protected Resource Metadata (RESOURCE-001)
RFC 9728 defines /.well-known/oauth-protected-resource which agents use to discover the authorization server. Without it, discovery falls back to fragile heuristics and spec-strict clients refuse to proceed. Only about 40% of public servers advertise it.
Tool-description poisoning (POISONING-001)
Tool descriptions containing zero-width unicode (U+200B, U+200C, U+200D) or bidirectional override characters (U+202A to U+202E) render as innocent names in UI but can encode instructions the LLM interprets. Real class of attack against LLM tool-selection.
Scan your server
npx -y @patchbay-nebius/agent audit <your-mcp-url>
Output is SARIF 2.1.0. Drop it in a GitHub Action, findings surface in Code Scanning next to your other SAST.
What is public
- Methodology: https://patchbay-mcp.dev/scanning
- Leaderboard of 277 servers: https://patchbay-mcp.dev/leaderboard
- MIT source: https://github.com/GodSmasher/patchbay-audit
Every finding has a tier label: source-confirmed, runtime-verified, or potential-fp, so you know what the scanner actually observed vs. inferred.
If you ship an MCP server and want it added to the public cohort, drop the URL in a comment.
