🛡️ A.I.G（AI-Infra-Guard）

中文版

A comprehensive, intelligent, easy-to-use, and lightweight AI Infrastructure Vulnerability Assessment and MCP Server Security Analysis Tool, developed by Tencent Zhuque Lab.

工具反馈问卷

A.I.G（AI-Infra-Guard）年初发布至今一直保持快速迭代中，为了打造更全面、智能、易用的全新的AI安全红队平台并吸引更多用户共建，在此特别邀请您参与一个5分钟的用户需求调研，对于有价值的反馈将我们后续会安排寄送腾讯特色小礼品，感谢大家的关注与支持。

https://doc.weixin.qq.com/forms/AJEAIQdfAAoAbUAxgbhAIQCNfZu1TQeRf

Table of Contents

• 🚀 Quick Preview
• ✨ Project Highlights
• 📋 Feature Description
• 🤝 MCP Security Certification and Cooperation
• 📦 Installation and Usage
• Installation
• Command Line Structure
• Usage
• WebUI Visual Operation
• Security Vulnerability Scanning (scan)
• MCP Server Scanning (mcp)
• 📊 Covered MCP Security Risks
• 📊 Covered AI Component Vulnerabilities
• 🔍 Component Fingerprinting Rules
• Example: Gradio Fingerprint Rule
• Fingerprint Matching Syntax
• 🤝 Contribution
• 📄 License

🚀 Quick Preview

MCP Server Code/Online Service Analysis
<br>
<img src="img/mcp-en.png" alt="MCP Server Code Detection Preview" height="600">
<br>

Infrastructure Vulnerability Scanning
<br>
<img src="img/scan-en.png" alt="Infrastructure Detection Preview" height="600">
<br>

✨ Project Highlights

• Comprehensive Security Assessment Capabilities
• Supports analysis of 9 common MCP security risks, continuously updated.
• Supports identification of 28 AI component frameworks, covering 200+ vulnerability fingerprints.
• Supports private deployment for easy integration into internal security scanning pipelines.
• Intelligent and User-Friendly Experience
• MCP security analysis driven by AI Agent for one-click intelligent analysis.
• AI component vulnerability scanning supports custom fingerprints and YAML vulnerability rules.
• Out-of-the-box usability with no complex configuration required, providing a Web interface for visual operation.
• Lightweight Design
• Core components are concise and efficient.
• Small binary size and low resource consumption.
• Cross-platform support (Windows/MacOS/Linux).

🤝 MCP Security Certification and Cooperation

AI Infra Guard is committed to providing professional MCP security analysis and certification solutions. We welcome MCP marketplaces, developer platforms, and hosting providers to integrate our tool into their pre-listing security scanning process for MCP Servers and display the scan results in the MCP marketplace, collectively building a safer MCP ecosystem.

If you are interested in cooperating with us, please contact Tencent Zhuque Lab at zhuque [at] tencent.com.

We also welcome you to share your implementation cases within the MCP community.

📋 Feature Description

AI Infra Guard consists of three core modules:

1. AI Component Vulnerability Scanning (scan): Detects known security vulnerabilities in web-based components within AI infrastructure.
2. MCP Security Analysis (mcp): Analyzes security risks in MCP Server code based on AI Agent.
3. WebUI Mode (webserver): Enables the web-based visual operation interface.

📦 Installation and Usage

Installation

Download the latest version suitable for your operating system from the Releases page.

Command Line Structure

AI Infra Guard uses a subcommand structure:

./ai-infra-guard <subcommand> [options]

Main subcommands:

• scan: Executes AI component security vulnerability scanning.
• mcp: Executes MCP Server code security analysis.
• webserver: Starts the Web interface server.

Usage

WebUI Visual Operation

Start the web server, listening on 127.0.0.1:8088 by default:

./ai-infra-guard webserver

Specify the listening address and port:

./ai-infra-guard webserver --ws-addr <IP>:<PORT>

Example: ./ai-infra-guard webserver --ws-addr 0.0.0.0:9090

AI Component Vulnerability Scanning (scan)

Local One-Click Detection (Scans common local service ports):

./ai-infra-guard scan --localscan

Scan a Single Target:

./ai-infra-guard scan --target <IP/Domain>

Example: ./ai-infra-guard scan --target example.com

Scan Multiple Targets:

./ai-infra-guard scan --target <IP/Domain1> --target <IP/Domain2>

Example: ./ai-infra-guard scan --target 192.168.1.1 --target example.org

Read Targets from a File:

./ai-infra-guard scan --file target.txt

The target.txt file should contain one target URL or IP address per line.

View Full Parameters for the scan Subcommand:

./ai-infra-guard scan --help

scan Subcommand Parameter Description:

Usage:
 ai-infra-guard scan [flags]

Flags:
 --ai Enable AI analysis (requires LLM Token configuration)
 --check-vul Validate the effectiveness of vulnerability templates
 --deepseek-token string DeepSeek API token (for --ai feature)
 -f, --file string File path containing target URLs
 --fps string Fingerprint template file or directory (default: "data/fingerprints")
 --header stringArray Custom HTTP request headers (can be specified multiple times, format: "Key:Value")
 -h, --help Show help information
 --hunyuan-token string Hunyuan API token (for --ai feature)
 --lang string Response language (zh/en, default: "zh")
 --limit int Maximum requests per second (default: 200)
 --list-vul List all available vulnerability templates
 --localscan Perform local one-click scan
 -o, --output string Result output file path (supports .txt, .json, .csv formats)
 --proxy-url string HTTP/SOCKS5 proxy server URL
 -t, --target stringArray Target URL (can be specified multiple times)
 --timeout int HTTP request timeout in seconds (default: 5)
 --vul string Vulnerability database directory (default: "data/vuln")

MCP Server Security Risk Analysis (mcp)

This feature automatically analyzes security issues in MCP Server code using an AI Agent.

Basic Usage (Uses OpenAI API by default, requires Token):

./ai-infra-guard mcp --code <source_code_path> --model <model_name> --token <api_token> [--base-url <api_base_url>]

Example: ./ai-infra-guard mcp --code /path/to/mcp/server --model gpt-4 --token sk-xxxxxx

Specify Output Format:

./ai-infra-guard mcp --code <source_code_path> --model <model_name> --token <api_token> --csv results.csv --json results.json

View Full Parameters for the mcp Subcommand:

./ai-infra-guard mcp --help

mcp Subcommand Parameter Description:

Usage:
 ai-infra-guard mcp [flags]

Flags:
 --base-url string LLM API base URL (optional, overrides default OpenAI URL)
 --code string Path to the MCP Server source code to scan (required)
 --csv string Output results to a CSV file path
 -h, --help Show help information
 --json string Output results to a JSON file path
 --log string Log file save path
 --model string AI model name (required, e.g., gpt-4, gpt-3.5-turbo)
 --plugins string Specify enabled plugins list (comma-separated, optional)
 --token string LLM API token (required)

📊 Covered MCP Security Risks

AI Infra Guard can analyze the following common MCP security risks, with continuous updates:

📊 Covered AI Component Vulnerabilities

AI Infra Guard supports detection of known vulnerabilities in various AI-related components:

Note: The vulnerability database is continuously updated.

🔍 Component Fingerprinting Rules

AI Infra Guard uses YAML-based rules for web component fingerprinting and vulnerability matching.

• Fingerprint Rules: Stored in the data/fingerprints directory.
• Vulnerability Rules: Stored in the data/vuln directory.

Example: Gradio Fingerprint Rule (data/fingerprints/gradio.yaml)

info:
 name: gradio
 author: Security Team
 severity: info
 metadata:
 product: gradio
 vendor: gradio
http:
 - method: GET
 path: '/'
 matchers:
 # Match if the response body contains Gradio-specific JavaScript configuration or elements
 - body="<script>window.gradio_config = {" || body="document.getElementsByTagName(\"gradio-app\");"

Fingerprint Matching Syntax

Match Locations

• title: HTML page title
• body: HTTP response body
• header: HTTP response header
• icon: Hash value (e.g., MurmurHash3) of the website's favicon

Logical Operators

• =: Fuzzy contains match (case-insensitive)
• ==: Exact equals match (case-sensitive)
• !=: Not equals match
• ~=: Regular expression match
• &&: Logical AND
• ||: Logical OR
• (): Used for grouping to change operator precedence

🤝 Contribution

We welcome community contributions!

• Report Issues: Submit an Issue
• Submit Code: Create a Pull Request

Wechat Group

📄 License

This project is open-sourced under the MIT License. For details, please refer to the License.txt file.