Samurai Security

What security tests do

What Samurai Security's live security test checks on your own MCP server, who can run it, what you need first, and how Guardian keeps it safe.

Samurai Security runs an authorized, live security test against one of your own MCP servers before you ship it. It looks for the ways a server facing AI agents you don't control is most often broken. The test is deterministic and uses no AI.

What it checks

Every run checks your server, and each of its tools, against seven attack classes:

  • Authentication bypass: calls each read-only tool with no credentials and reports any call the server should have refused.
  • Token forgery: calls each read-only tool with your saved credential damaged into an inert value and checks that the server rejects it.
  • OAuth misconfiguration: checks that an unauthorized request gets a proper authentication challenge, and that the URLs in your OAuth metadata are safe and match where they are published.
  • DNS rebinding: sends a request with a foreign Origin header and checks that the server rejects it, as the MCP transport spec requires.
  • Secret exposure: scans real tool responses for API keys, tokens and private keys. Only the secret type and a short prefix are kept.
  • Prompt injection: scans real tool responses for text that tells an AI agent to ignore its instructions or leak credentials.
  • Path traversal: sends a ../ path to tools that take a file or path and checks that no system file comes back.

A check that can't be tried, for example because the server has no OAuth metadata or a tool has no path-like argument, is shown as Unverified with the reason, not as a pass.

Who can run it

Only Admins and Security members can run a security test, and only on a server whose credentials they saved themselves. Everyone else on the team, whatever their role, can open the Security tests page to read the history and results.

Before you can run one

  • Your plan includes Samurai Security. It is part of the Standard and Plus plans.
  • An Admin has accepted the Security Terms. An Admin accepts them once on the Security tests page, and they cover everyone on the team. They state that live requests are sent to the server you select, that they include a path-traversal attack payload, that calls use the saved server's configuration (its credential headers, if any), and that your security team approves and supervises every run, including runs from the Guardian CLI. Every report records who authorized it and when.
  • The server is saved, and is an http server. You can only test a server saved on the MCP Servers page, never a URL you type in. A stdio server can't be tested.
  • For the web app, the server is marked production. The web app lists only production servers. The Guardian CLI doesn't have this requirement.

How it stays safe

  • Read-only tool calls. A run calls only tools marked read-only. Other tools are skipped and listed as Skipped tools, and never called.
  • Bounded work. A run tests a limited number of tools, sends a limited number of requests one at a time, and stops at a time limit. A check that runs out of time is shown as unverified.
  • One run per server. A second test on the same server waits until the first has finished.
  • Identified traffic. Every request Guardian sends carries a User-Agent: Guardian/... header and a Guardian-Run-Id header, so the owner of a server can tell Guardian's traffic from anything else and report it with the run id. Server owners can read how at the public /abuse page, which needs no sign-in.
  • No stored secrets. Results keep only fixed outcomes, such as a secret's type and short prefix, never the raw response.

Only run a test against servers you own or are authorized to test, ideally a non-production environment.

Next steps

To run a test, see Run security tests.