An MCP server gives an AI agent a set of tools, such as "add two numbers" or "fetch a report". MCP security is the work of making sure each tool does only what the calling user may do.
We built a small MCP server with a vulnerable mode and a fixed mode, and wrote tests for three flaws. We ran them on 12 November 2026 and again on 13 November 2026, and both runs gave the same 16 results. In vulnerable mode all 4 attack checks succeeded. In fixed mode all 4 were blocked. The 3 control checks gave the same answer in both modes, so the fixes did not break normal use.
The three flaws are a tool description that carries hidden instructions, a user token passed on to another API, and a tool that checks who you are but not what is yours. Each can leak data, and each has a short test. The server, tests and results are in the mcp-security-fixture repository (commit 194b4a6). For each flaw this page describes the flaw, what the test checks and the fix.…