Scanning AI infrastructure like it’s a network pentest target

What it does AI-Infra-Guard is a full-stack red-teaming platform from Tencent’s Zhuque Lab. It fingerprints live AI services—Ollama, vLLM, ComfyUI, n8n, and over 60 others—to match them against a database of more than 1,300 known CVEs. Beyond infrastructure, it scans AI agent workflows, MCP servers, and agent skills for 14 categories of security risks, and runs jailbreak evaluations against LLMs using curated attack datasets.

The interesting bit Most security tools treat the model as the only target; A.I.G treats the entire surrounding stack as the target. Its Agent Scan spins up a multi-agent framework to autonomously probe workflows on platforms like Dify and Coze, while the MCP/Skills scanner can audit both local source code and remote URLs. It is essentially a vulnerability scanner that learned where AI infrastructure lives.

Key highlights

• Fingerprints over 64 AI components (Ollama, vLLM, ComfyUI, n8n, Triton, etc.) against more than 1,300 CVEs and known misconfigurations.
• Scans MCP servers and agent skills across 14 risk categories from source or remote URLs.
• Multi-agent automated framework for probing AI agent workflows on platforms like Dify and Coze.
• Jailbreak evaluation with curated datasets and cross-model comparison capabilities.
• Ships as a Dockerized web service with a complete API and Swagger documentation.

Caveats

• Explicitly positioned for internal use only: it currently lacks authentication and the README warns against deploying it on public networks.

Verdict Worth a look if you run internal AI infrastructure and want to discover CVEs and misconfigurations before attackers do. Skip it if you need a hardened, public-facing scanner or if your threat model stops at the prompt layer.