Skip to content

SECURITY.md

Latest commit

 

History

History
55 lines (41 loc) · 2.16 KB

SECURITY.md

File metadata and controls

55 lines (41 loc) · 2.16 KB
Raw

Security policy

Supported versions

Version Supported
0.1.x Yes

Reporting a vulnerability

If you find a security issue in mcp-fence itself (the scanner, fuzzer, sandbox runner, or report generator):

  1. Do not open a public GitHub issue.
  2. Email the maintainers, or open a GitHub security advisory via the "Security" tab on the project repository.
  3. Include reproduction steps, affected version, and impact assessment.

We aim to acknowledge reports within 5 business days.

Scope: what mcp-fence is and is not

mcp-fence is a defensive local tool used by MCP server authors and operators to audit their own servers. It is not an offensive tool and is not intended for unauthorized testing of third-party services.

Safety guarantees

  • Non-destructive. Fuzzing payloads are harmless; the marker payload used to detect command injection is echo MCPG_FUZZ_MARKER_8f2a. No rm, mv, chmod, or destructive system calls are ever emitted.
  • Local-first. No code, configuration, or scan result is ever uploaded. The optional LLM judge talks to a local endpoint only (Ollama, vLLM, or any OpenAI-compatible local server).
  • No public-network scanning. SSRF payloads test the URL validation behaviour of the target tool. The default fuzz runner does not initiate outbound HTTP requests itself.
  • Safe path probing. Path-traversal fuzzing aims at fake secret files inside the scanned project. It never targets /etc/shadow, ~/.ssh/id_*, or other real sensitive paths.

What you opt into with --allow-unsafe

--allow-unsafe lets the fuzzer pass through the unsafe-tool gate so it can call tools whose name or argument schema looks shell-like (command, shell, exec, code, script, ...). This is appropriate inside the provided Docker sandbox or against your own toy servers; it should not be used against unfamiliar third-party MCP servers running on your machine.

Use of examples

The examples/vulnerable_* servers in this repository contain intentionally insecure code for testing. Do not run them outside the sandbox or expose them on a network interface.