Security at Guardian Samurai

Guardian Samurai tests public and private MCP servers, so we are very careful with all the information given to us and what we connect to. This page describes the controls that are in place today.

How Guardian Samurai connects to your MCP servers

You can test an MCP server two ways: from the Guardian Samurai web app, where Guardian Samurai connects to your server, or with the guardian CLI, where your own machine connects and Guardian Samurai never reaches the server.

The diagram shows what data moves in each case, where it is stored, and the safeguards along the way.

Web app pathCLI pathSafeguard
YOUR ENVIRONMENTGUARDIAN SAMURAI CLOUDINTERNET · HTTPSYour browserGuardian Samurai web appYour MCP serverthe server being testedguardian CLIyour laptop or CIlocal-only by defaultyou confirm each runGuardian Samurai APIsign-in requiredeach customer isolatedTest runnerweb app runs onlypublic addresses onlyread-only tools for scanssize and time limitsCLI plans and uploadsscoped CLI tokenuploads must match planCredential vaultsaved server secretsAWS KMS · AES-256Results storereports and historysecrets scrubbedencrypted at restW1W5W2W3W4C1 · C3C2C4C4

Web app path

  1. W1You pick a saved server and start a scan, a functional test or a Samurai Security test.
  2. W2Guardian Samurai decrypts that server's saved credentials, in memory, for this run only.
  3. W3Guardian Samurai's test runner connects to your MCP server over the internet and runs the checks.
  4. W4Guardian Samurai strips secrets from the report and stores it encrypted.
  5. W5You read the report in the web app.

CLI path

  1. C1The CLI connects to your MCP server from your machine and reads its list of tools.
  2. C2It sends Guardian Samurai a description of those tools, with secrets masked, and gets a test plan back.
  3. C3After you confirm, it runs the plan against your server, locally.
  4. C4It uploads the outcomes, if you allow it. Results always print locally too.

How we handle your secrets

  • Run configs are discarded unless you save the server. Configs, headers, environment values and raw responses from your server are used during the run and then discarded. Findings name the type of secret we found (for example "API key in env"), never the value.
  • Saved servers are encrypted. When you're signed in, MCP Scan can save a server so you don't have to paste it again. The "Save server" box is ticked by default; untick it and nothing is kept. A saved server's URL, headers, and stdio command, arguments and environment values are encrypted with AES-256-GCM under a fresh data key from AWS KMS, bound to your organization and to that server. Only the server's name and its address with credentials removed are stored unencrypted. A saved secret is never shown again: editing shows •••••••• in its place. Any member of your organization can scan, edit or delete its saved servers, and deleting one removes it for good.
  • Reports are scrubbed before they're saved. Headers, env, args and config are removed, and credentials and query strings are stripped from URLs. Evidence must be masked or the report isn't written.
  • Anonymous scans are never saved. If you're not signed in, the scan runs and the results are returned to you, and that's all.
  • Request bodies are never logged. Our logs record only allow-listed metadata.

When Guardian Samurai connects to your server

  • We never run your stdio servers in the cloud. Cloud scans only contact remote (HTTP) servers.
  • We refuse to connect to internal addresses. Before connecting, we resolve the hostname and refuse the target if any of its addresses is private, loopback, link-local (including cloud metadata addresses) or reserved. We then connect to the address we checked, so DNS tricks can't redirect us. We don't follow redirects.
  • Every connection has hard limits. Each request has a fixed deadline and a response size cap, and each run has its own deadline, so a slow or hostile server can't tie us up.
  • Our requests are identifiable. Every request Guardian Samurai's cloud sends carries a Guardian User-Agent and a Guardian-Run-Id. If you own a server and see traffic you didn't expect, the Report Guardian traffic page explains how to recognize and report it.
  • We don't send your credentials to identity probes. Probes send no credentials and keep only derived yes/no results.

The Guardian CLI (local testing)

The CLI runs on your machine. Safety decisions are made on your machine, not by our cloud:

  • Target allow-list. The CLI only connects to loopback addresses unless you name a host with --allow-target, and it pins the address it checked.
  • Preview before anything runs. You see which tools will be called and which were skipped, and you confirm. Only an explicit --yes flag skips the prompt. No config file or environment variable can skip it.
  • Side-effect gate. Tools that can write, delete or send are skipped unless you allow them on the command line. A cloud plan can't unlock them. A tool with no annotations runs only if its name is on a published read-only list.
  • One small binary with no third-party code. The CLI is a single Go binary built only from the Go standard library, and CI fails the build if a dependency is added.
  • Security checks on every change. CI runs staticcheck, gosec and govulncheck, plus tests against a deliberately hostile test MCP server.

Accounts and access

  • Passwords are hashed with scrypt and a per-user salt. Sign-in with Google is also supported.
  • Sessions use random tokens that we store only as hashes. They are held in HttpOnly, SameSite=Strict, Secure cookies and expire after 12 hours idle or 7 days at most. Resetting your password signs you out everywhere.
  • No account enumeration. Sign-in, signup and password reset give the same answer whether or not an account exists.
  • Rate limits apply to sign-in, signup, password reset and scan submission.

Keeping customers separate

  • Every customer's data is isolated with Postgres row-level security, and the application also filters by customer explicitly. Queries that run without a customer context see no rows.
  • IDs are random UUIDs. Another customer's ID looks exactly like one that doesn't exist.

Staff access

  • Staff sign in to a separate admin console with a password and a second factor every time, and any change to a customer account requires a fresh second-factor check. Every change is recorded and triggers an alert.
  • Staff can't sign in as you, can't see your scan or test results, and can't read your saved servers' secrets, which are stored only in encrypted form.
  • Access to production servers and the database is limited to a small number of people, requires MFA, and is logged. No shared or long-lived keys are used.

What we don't have yet

  • No SOC 2 or ISO 27001 report yet.
  • No third-party penetration test yet.
  • Customer accounts don't support MFA yet. If you need MFA today, use Sign in with Google with your Google account's 2-step verification.

Reporting a vulnerability

Email security@guardiansamur.ai. We'll confirm receipt within 24 hours. Please give us reasonable time to fix the issue before disclosing it. See also /.well-known/security.txt. To report unexpected traffic from Guardian Samurai to a server you own, see Report Guardian traffic.