AI-Infra-Guard

Overview

AI-Infra-Guard is a full-stack AI Red Teaming platform developed by Tencent Zhuque Lab, designed to secure AI ecosystems through automated vulnerability scanning across agent infrastructure, MCP servers, agent skills, and LLM models. Released under the Apache 2.0 license, it has rapidly gained traction, reaching approximately 5,200 GitHub stars.

The platform emerged in response to the growing attack surface of AI agent systems — which combine traditional software vulnerabilities with novel agent-specific risks like prompt injection, tool misuse, and unauthorized data exfiltration. AI-Infra-Guard addresses these threats through a unified scanning pipeline that covers the full stack of an AI deployment.

The project is primarily written in Go for its backend, with Python and JavaScript/TypeScript components for CLI tools, API checkers, and the web frontend.

Features

• AI Infrastructure Vulnerability Scan: Identifies 100+ AI framework components and 2,000+ known CVEs across Ollama, ComfyUI, vLLM, n8n, and Triton Inference Server
• MCP Server & Agent Skills Scan: Detects 14+ categories of security risks; supports scanning from source code or remote URLs
• Agent Scan: Multi-agent automated framework for evaluating AI agent workflow security, including OWASP skills and web-exfiltration detection
• LLM Jailbreak Evaluation: Assesses prompt security with Many-Shot, PAIR, GOAT, and ActorAttack methods with cross-model comparison
• ClawScan: One-click evaluation of OpenClaw security risks including insecure configurations and CVE vulnerabilities
• aig-skill-scan CLI: Pip-installable CLI for CI/CD integration, classifying risks per SkillTrustBench T01–T09 taxonomy
• Model & API Relay Checker: Model fingerprinting, Claude Signature verification, and relay auditing
• Web UI: User-friendly interface at localhost:8088 with one-click scanning and real-time progress
• Complete REST API: Full Swagger documentation for enterprise integration

Installation

Full Platform

git clone https://github.com/Tencent/AI-Infra-Guard.git
cd AI-Infra-Guard

# Build with Go
go build -o aig-guard ./cmd/server

# Run the server
./aig-guard --port 8088

CLI Tool Only

pip install aig-skill-scan

Quick Start

Skill Scan via CLI

# Scan a local skill directory
aig-skill-scan ./skills/

# Scan with CI/CD integration
aig-skill-scan ./skills/ --format json --output report.json

MCP Server Scan

// Via REST API
POST /api/mcp/scan
{
  "target": "https://example.com/mcp-server",
  "depth": "full"
}

Core Concepts

Defense in Depth for AI

AI-Infra-Guard follows a defense-in-depth model, scanning at each layer of the AI stack:

1. Infrastructure Layer: AI tools and frameworks (Ollama, vLLM, n8n)
2. Protocol Layer: MCP servers and their tool definitions
3. Agent Layer: Agent skills, prompts, and workflows
4. Model Layer: LLM jailbreak susceptibility and model integrity

SkillTrustBench Taxonomy

The CLI tool classifies risks using the SkillTrustBench T01–T09 taxonomy, which covers:

• T01: Sensitive file access
• T02: Network exfiltration
• T03: Code execution
• T04: Prompt injection
• T05: Tool misuse
• T06: Memory pollution
• T07: Configuration weakness
• T08: Authentication bypass
• T09: Logging vulnerabilities

Advanced Features

CI/CD Integration

Integrate skill scanning into your CI/CD pipeline to catch vulnerabilities before deployment:

# .github/workflows/ai-security.yml
name: AI Skill Security
on: [push, pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Scan AI skills
        run: |
          pip install aig-skill-scan
          aig-skill-scan ./skills/ --fail-on high

Model Fingerprinting

Identify which model an API is actually running, preventing model substitution attacks:

aig-skill-scan --check-model https://api.example.com/v1/chat/completions
# Output: Detected model: gpt-4o-2026-05 (confidence: 94%)

Agent Workflow Security Test

Spawn automated agent probes that attempt to exploit agent workflows for data exfiltration, privilege escalation, or unauthorized tool usage.

Examples

Scanning an n8n Deployment

aig-skill-scan --target http://localhost:5678 --type n8n

Cross-Model Jailbreak Evaluation

POST /api/jailbreak/evaluate
{
  "prompts": ["prompt_injection_attempt_1"],
  "models": ["gpt-4o", "claude-opus-4-7", "deepseek-v4-flash"],
  "methods": ["PAIR", "GOAT"]
}

Pros

• Comprehensive coverage across AI infrastructure, MCP, agents, and models
• Apache 2.0 license allows commercial use without restrictions
• Tencent Zhuque Lab backing ensures active development and updates
• CI/CD-ready CLI tool for automated vulnerability scanning
• Web UI accessible to non-technical security teams
• SkillTrustBench taxonomy aligns with OWASP Agentic Top 10

Cons

• New project with limited production track record
• Go backend may require specific deployment environment
• Automated scans may produce false positives requiring manual triage
• Primarily focused on open-source AI tools — proprietary platforms may not be covered
• Model fingerprinting accuracy depends on API response characteristics

When to Use

AI-Infra-Guard is essential for:

• Organizations deploying AI agents in production that need automated security scanning
• Teams using MCP servers who need to verify tool security before integration
• DevSecOps pipelines that need CI/CD-integrated AI vulnerability scanning
• Security teams evaluating agent skills from untrusted sources
• Organizations subject to regulatory scrutiny around AI security (e.g., EU AI Act)

Resources

• GitHub Repository
• Web UI (when running locally)
• Tencent Zhuque Lab
• OWASP Agentic Top 10
• SkillTrustBench