mcpsec

mcpsec

Security scanner and protocol fuzzer for MCP servers

Installation • Quick Start • Audit v3 • Scanners • Fuzzing

---

Why mcpsec?

MCP (Model Context Protocol) connects AI agents to external tools. Claude Desktop, Cursor, VS Code Copilot, and every major AI IDE uses it. Security is often an afterthought.

Most MCP security tools do static analysis. mcpsec connects to live servers and proves exploitation.

---

Real Bugs Found

Target	Vulnerability	Severity	Status
mobile-mcp	URL Scheme Injection (CVE-2026-35394, CVSS 8.3) - Arbitrary code execution via unsanitized tool input	High	Fixed - PR #299
MCP Python SDK	ClosedResourceError DoS (invalid UTF-8)	High	Issue #2328 - Fix in PR #2334
radare2-mcp	Arbitrary RCE via shell escape (!) in run_command/run_javascript	Critical	Issue #45 - Fixed in commit 482cde6
radare2-mcp	Multiple SIGSEGV via params type confusion	High	Issue #42
radare2-mcp	SIGSEGV in initialize via params type confusion	High	Issue #52
MCP Python SDK	UnicodeDecodeError DoS	Medium	Fixed - PR #2302
mcp-server-fetch	61 crash cases, exception handling DoS	High	Issue #3359
mcp-server-git	61 crash cases	High	Issue #3359
MCP TypeScript SDK	EPIPE crash	Medium	Issue #1564
MCP TypeScript SDK	Integer overflow DoS (MAX_SAFE_INTEGER+1)	Medium	Issue #1765

More findings under responsible disclosure.

---

Installation

pip install mcpsec

For AI-powered features:

pip install mcpsec[ai]

Nix

nix-shell   # basic
nix-shell --arg withAll true   # all optional deps

---

Quick Start

Runtime Scanning

# Scan via stdio
mcpsec scan --stdio "npx @modelcontextprotocol/server-filesystem /tmp"

# Scan via HTTP with auth
mcpsec scan --http http://localhost:8080/mcp -H "Authorization: Bearer TOKEN"

# Auto-discover and scan all local servers
mcpsec scan --auto

# Enumerate attack surface
mcpsec info --stdio "python my_server.py"

Static Analysis (Audit v3)

# Local source - pattern-based + AI reachability
mcpsec audit --path ./my-mcp-server

# GitHub repository
mcpsec audit --github https://github.com/user/mcp-server

# With LLM-powered taint analysis
mcpsec audit --github https://github.com/user/mcp-server --ai

# Known vulnerable servers
mcpsec audit --github https://github.com/radareorg/radare2-mcp

Protocol Fuzzing

# Standard fuzzing (~200 cases)
mcpsec fuzz --stdio "python my_server.py"

# High intensity (~800 cases)
mcpsec fuzz --stdio "python my_server.py" --intensity high

# AI-powered payload generation
mcpsec fuzz --stdio "python my_server.py" --ai

Advanced

# SQL Injection scanner with DB fingerprinting
mcpsec sql --stdio "npx @benborla29/mcp-server-mysql" --fingerprint

# Dangerous tool chain detection
mcpsec chains --stdio "npx @example/complex-server"

# Interactive exploitation REPL
mcpsec exploit --stdio "npx vulnerable-server"

# Rogue server for client-side testing
mcpsec rogue-server --port 9999 --attack all

---

Static Analysis - Audit v3

New in v2.7.1 - Complete rewrite of the audit engine with a pattern-based architecture.

7-Stage Analysis Pipeline

Source Code
    │
    ├─ 1. Fetch        - Clone GitHub repo or load local path
    ├─ 2. Detect       - Identify language, MCP SDK, and framework
    ├─ 3. Sink Scan    - 3,450+ regex patterns across 12 languages
    ├─ 4. Semgrep      - 149 semantic rules (AST-level)
    ├─ 5. AST          - Python/JS taint flow analysis
    ├─ 6. Reachability - LLM taint tracing (heuristic fallback)
    └─ 7. Deduplicate  - Merge, rank, and report findings

Pattern Database - 3,450+ Sink Patterns

Vulnerability Class	Patterns	Languages
Command Injection	181	Python, JS/TS, Go, Rust, Java, C, C#, Ruby, PHP
SQL / NoSQL Injection	~100	All drivers + ORM-specific (Sequelize, SQLAlchemy, Drizzle, Kysely)
Path Traversal	~60	fs, aiofiles, Deno, Bun, tarfile, ZipSlip
SSRF	~80	requests, httpx, aiohttp, gRPC, OkHttp, WebSocket, got
Deserialization	~60	pickle, YAML, torch.load, numpy, joblib, BinaryFormatter
Code Execution	~50	eval, vm, exec, DOM XSS, format strings
Template Injection	~30	Jinja2, Pug, EJS, Handlebars, Lodash, ERB, Velocity, Thymeleaf
Crypto Weaknesses	~40	MD5/SHA-1, RC4, weak keys, JWT none alg
XXE	~25	lxml, untangle, DOMDocument, SAXParser
Log/Header/LDAP	~50	All major frameworks
Prototype Pollution	~15	Object.assign, deepmerge, __proto__
Sanitizers	105	Command, SQL, Path, XSS sanitizers (Python, JS, Go, Rust)
MCP-Specific	~45	Tool args → sinks, prompt/resource handlers (20+ SDKs)

Framework Detection

Automatically identifies:

• MCP SDKs: @modelcontextprotocol/sdk, mcp (Python), mcp-go, rmcp (Rust), mcpx (C#)
• Languages: TypeScript, JavaScript, Python, Go, Rust, Java, C#, PHP, Ruby, C/C++
• Frameworks: Express, FastAPI, Django, Gin, Axum, Spring Boot, ASP.NET

Heuristic Fallback

When no LLM is configured, the reachability analyzer uses confidence scoring to report findings without false silence - high-confidence patterns (CRITICAL/HIGH + direct taint) are always reported.

---

Scanners

Scanner	Description
prompt-injection	Hidden instructions in tool descriptions
command-injection	OS command injection with 138 payloads
path-traversal	Directory traversal with 104 payloads
ssrf	Server-Side Request Forgery with 81 payloads
sql	SQL Injection (Error, Time, Boolean, Stacked)
auth-audit	Missing authentication, dangerous tool combos
description-prompt-injection	LLM manipulation via descriptions
resource-ssrf	SSRF via MCP resource URIs
capability-escalation	Undeclared capability abuse
chains	Dangerous tool combination detection
code-execution	Detects eval(), exec(), and compile() sinks
template-injection	Targets SSTI and string formatting vulnerabilities
rag-poisoning	Identifies dangerous Write→Read data flows
idor	Insecure Direct Object Reference detection
info-leak	Environment variable and credential disclosure
deserialization	Pickle, XXE, and unsafe YAML parsing

---

Fuzz Generators

22 generators organized by intensity level:

Low (~65 cases): malformed_json, protocol_violation, type_confusion, boundary_testing, unicode_attacks

Medium (~200 cases): + session_attacks, encoding_attacks, integer_boundaries

High (~800 cases): + injection_payloads, method_mutations, param_mutations, timing_attacks, header_mutations, json_edge_cases, protocol_state, id_confusion, concurrency_attacks, regex_dos, deserialization

Insane (~1500+ cases): + resource_exhaustion, memory_exhaustion

---

How It Works

┌─────────┐     MCP Protocol      ┌────────────┐
│ mcpsec  │ ◄──── JSON-RPC ────►  │   Target   │
│         │    (stdio / HTTP)     │   Server   │
└────┬────┘                       └────────────┘
     │
     ├── Connect & enumerate attack surface
     ├── Run 10+ security scanners
     ├── Generate 800+ fuzz cases
     ├── Execute AI-powered payload mutations
     ├── Static audit: 3,450+ sink patterns + 149 Semgrep rules
     └── Report findings with PoC evidence

---

Configuration

AI Provider Setup

mcpsec setup

Supports: OpenAI, Anthropic, Google, Groq, DeepSeek, Ollama

Output Formats

# JSON
mcpsec scan --stdio "server" --output results.json

# SARIF 2.1.0 (GitHub/GitLab/Azure DevOps CI/CD)
mcpsec fuzz --stdio "server" --output results.sarif

---

See CHANGELOG.md for the full release history.

---

Contributing

See CONTRIBUTING.md for guidelines.

CI runs automatically on all PRs - linting with Ruff and cross-platform tests (Ubuntu, Windows, macOS).

---

Disclaimer

For authorized security testing only. Only scan servers you own or have explicit permission to test.

---

License

MIT

---

Built by Manthan Ghasadiya