convez.ai
MCP Server
Den-Sec

mcprobe

Den-Sec/mcprobe

Ver no GitHub
InicialPluginScore · de 0 a 100

Active, confirmation-driven security scanner for MCP servers - probes tool handlers for command injection, SSRF, path traversal, auth bypass & secret leaks and confirms them out-of-band. Burp Active Scan, for MCP.

Instalação

$npx convez-ai Den-Sec/mcprobe

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
v0.5.0
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

mcprobe

Active, confirmation-driven security scanner for MCP server implementations - Burp Active Scan, for MCP.

mcprobe enumerates the tools exposed by a Model Context Protocol (MCP) server, maps their parameters to injection points, fires targeted payloads, and reports each finding with a confidence level earned by a concrete oracle - prioritising actively-confirmed, exploitable issues over guesses.

Why mcprobe

Most "MCP security" tooling is either a generic fuzzer (noisy, low-signal) or a defensive/static analyzer (looks at config and source, never proves exploitability). mcprobe is different:

  • Oracle-backed, not guesses. Every finding is tied to a concrete signal: an out-of-band (OOB) callback, a calibrated time delay, a canary value reflected in the response, or a baseline diff. No signal, no finding.
  • Graded, calibrated confidence. Findings carry an explicit confidence level (CONFIRMED / FIRM / TENTATIVE, see Confidence levels). Per-tool baseline calibration suppresses the usual false-positive classes - a slow-but-safe tool, or output that merely looks secret-shaped, is not flagged.
  • Both transports. Works against MCP servers over stdio (local process) and streamable HTTP (remote endpoint, with custom headers/auth).

Install

pip install -e .

This installs the mcprobe console entry point.

Usage

Scan a local stdio server:

mcprobe scan --stdio "python server.py"

Scan a remote HTTP MCP endpoint, with auth, emitting SARIF:

mcprobe scan --http https://host/mcp --header "Authorization: Bearer X" --output sarif

Useful flags:

  • --stdio "<cmd>" launch the server as a subprocess and scan over stdio.
  • --http <url> scan a streamable HTTP MCP endpoint.
  • --header "k:v" add an HTTP header (repeatable).
  • --oob {local,interactsh,none} confirmation backend for OOB callbacks (local default; interactsh requires an injectable interactsh client, see below).
  • --aggressive also send blocking time-based (sleep) probes; by default mcprobe sends only non-blocking OOB/canary/pattern probes (time-based command-injection detection is aggressive-only).
  • --output {console,json,sarif,md} output format (default console).

Out-of-band (OOB) confirmation

OOB callbacks are how mcprobe confirms blind command injection and SSRF: a probe makes the target reach back to a listener mcprobe controls.

  • --oob local (default) spins up an in-process HTTP listener on localhost. It needs no external service and works for targets that can reach your machine (typically local stdio servers).
  • --oob interactsh uses an out-of-band interaction server for targets that cannot reach localhost (e.g. remote HTTP servers). mcprobe's InteractshOOB is a thin, client-agnostic wrapper: it expects an injectable client object exposing register() -> domain and poll() -> list. You supply that client; any library implementing those two methods works. If no such client is installed, --oob interactsh errors gracefully and tells you to use --oob local instead. No specific pip package is bundled or required.
  • --oob none disables OOB confirmation; only time-based and canary oracles run.

Confidence levels

Every finding carries one of three confidence levels, each earned by a specific oracle:

LevelMeaningHow it's earned
CONFIRMEDThe payload provably executed, or protected data was reached.An out-of-band callback fired (cmd injection, SSRF), a canary value was read back (path traversal), or an unauthenticated call returned the protected data (auth bypass).
FIRMA calibrated signal strongly indicates the issue, short of an OOB proof.A response delay exceeds the per-tool calibrated baseline by the injected sleep (time-based cmd injection), or a secret-shaped string appears in the probe response but not in the benign baseline (info leak).
TENTATIVEPattern-only match, with no calibration to corroborate it.A secret-shaped string matched, but no baseline was available to prove the input triggered it - review manually.

The OOB, canary, and auth-differential checks emit only CONFIRMED; the timing and info-leak oracles are where FIRM and TENTATIVE arise.

Checks (v1)

CheckVulnerabilityCWE
cmd_injectionOS command injectionCWE-78
ssrfServer-side request forgeryCWE-918
path_traversalPath traversalCWE-22
auth_bypassMissing authenticationCWE-306
info_leakSecret / sensitive info leakCWE-200

Authorized testing only

mcprobe is an active scanner. It sends real, potentially destructive payloads to the target. Run it only against systems you own or have explicit written authorization to test. Unauthorized use may be illegal. You are responsible for how you use this tool.

Validation

mcprobe is validated by an automated test suite against bundled deliberately-vulnerable fixture servers in tests/fixtures/. The suite exercises command injection (including cross-OS cmd.exe / PowerShell payloads), SSRF, path traversal, info-leak, nested/array/enum injection points, and the OOB, baseline-calibration, and false-positive-suppression paths end to end. See docs/claims-matrix.md for the claim-to-test mapping.

Roadmap

  • SQL injection check.
  • MCP-specific checks: tool-poisoning / prompt-injection via tool descriptions, and tool-scope / permission-boundary violations.
  • Additional OOB providers and richer time-based oracles.

License

MIT - see LICENSE.

Ver o original no GitHub

Relacionados