How to Check the Security of AI Services and Agents Using AI Infra Guard

Recently came across the AI-Infra-Guard repository (abbreviated as A.I.G) from Tencent's Zhuque Lab research team. The project addresses a specific and quite painful problem: as soon as you spin up local AI services like vLLM, Ollama, or ComfyUI, or start connecting third-party plugins and MCP servers to your LLM, the security perimeter instantly expands.

Standard vulnerability scanners poorly understand the specifics of the AI stack. They cannot test models for jailbreaks, do not check for instruction manipulation in agent skills, and are unaware of inference framework-specific signatures. The Tencent developers released a ready-made platform for continuous monitoring and automated auditing of the entire AI pipeline into open source.

What the scanner can do

The platform covers multiple security layers, from network infrastructure to prompt and tool logic.

1. Network-based AI infrastructure scanning

The tool connects to a running service via IP or URL and captures its digital fingerprint. The database recognizes over 100 components (Ollama, ComfyUI, vLLM, n8n, Triton Inference Server) and cross-references them against a database of more than 20,000 known CVEs.

You can scan a single instance or an entire subnet.

2. MCP server and agent skill auditing

With the spread of Model Context Protocol (MCP) and agent engines, a new threat vector emerged: malicious code inside pluggable tools, tool poisoning, and credential leakage.

A.I.G can analyze both the source code of an MCP server repository and uploaded archives. The scanner checks nine risk categories according to the SkillTrustBench classification:

• Skill instruction hijacking and context memory poisoning
• Covert executable code injection and payload downloads
• Privilege escalation in the system
• Insecure dependencies and tool spoofing

3. Jailbreak resistance assessment

The Jailbreak Evaluation module includes various LLM attack techniques (including multi-step attacks like Many-Shot, PAIR, GOAT, and ActorAttack). You connect the target model's endpoint via API, select a benchmark, and the scanner runs tests, producing a report with successful attack percentages and the model's weak spots.

Quick start

The simplest way to deploy the platform locally is to launch the ready-made image via Docker Compose. You will need a machine with 4 GB of RAM and about 10 GB of disk space.

git clone https://github.com/Tencent/AI-Infra-Guard.git
cd AI-Infra-Guard
docker compose -f docker-compose.images.yml up -d

The web interface will be available at http://localhost:8088.

Note: the repository explicitly states that the built-in server currently lacks an authentication mechanism. Exposing it to the public internet is strongly discouraged—run it only within an internal network or on a local machine.

Running skill audits via CLI

If you need to integrate agent checks into CI/CD without spinning up the web interface, the project has a separate CLI package:

pip install aig-skill-scan

export LLM_API_KEY="your-api-key"

# Сканирование директории с кодом скилла
aig-skill-scan --repo /path/to/your/skill \
               -m deepseek-v4-flash \
               --language en \
               -o result.json

How auditing works in practice

Suppose a vLLM inference instance is running on your local server on port 8000. To check it for known vulnerabilities:

1. Open the web interface at http://localhost:8088.
2. Navigate to the AI Infra Scan section.
3. Enter the network address http://127.0.0.1:8000 in the address field.
4. Run the check. The scanner will identify the component version, match it against the rules database, and display a list of discovered CVEs with links to patches.

Tool repositories are checked similarly: you provide a link to a GitHub MCP server implementation, and the scanner parses the syntax, checks permissions, path handling, and network calls.

Who will find this useful

The project targets teams that actively integrate LLMs and AI agents into their workflows:

• AppSec and security specialists: a ready-made ruleset for testing your company's AI stack without having to manually assemble exploits.
• MCP and agent developers: the CLI utility aig-skill-scan easily integrates into pre-commit hooks or GitHub Actions pipelines to check pluggable plugins before deployment.
• MLOps engineers: a quick way to scan your internal network and ensure no colleagues have exposed an unpatched Ollama or ComfyUI server externally.