One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

Tencent/AI-Infra-Guard

4,744+28 todayPython

Tencent open-sources a security scanner that checks AI services, agents, and MCP servers all at once

AI-Infra-Guard (A.I.G) is an AI red teaming platform from Tencent Zhuque Lab that scans AI infrastructure for known vulnerabilities, evaluates agent and MCP server risks, checks agent skills, and tests LLMs against jailbreak attacks. It can run as a Docker web app or as separate CLI tools plugged into CI/CD pipelines. It's released as open source under Apache 2.0.

What it does

  1. Point it at a running AI service address (vLLM, Ollama, ComfyUI, n8n, etc.) and it fingerprints the software to match against over 100 supported frameworks and more than 2,000 known CVE vulnerabilities
  2. MCP servers and Agent Skills can be scanned from source code or a remote URL alone, covering 14 major risk categories; on the project's own SkillTrustBench benchmark spanning 9 risk types (T01-T09), the top-performing model reached an F1 score of 0.9848
  3. Jailbreak Evaluation tests LLM safety using multiple attack methods, including multi-turn attacks like Many-Shot, PAIR, GOAT, and ActorAttack that try to manipulate a model over several conversation turns
  4. Skill, MCP, and Agent scanning are each packaged as standalone pip-installable CLI tools so they can be embedded directly into enterprise CI/CD pipelines
  5. The maintainers explicitly warn the tool has no authentication mechanism and is meant for internal enterprise or individual use, not public-facing deployment

Why it matters

Teams running their own AI services or building agent/MCP-based workflows get a free way to self-check for known vulnerabilities and permission-abuse risks before shipping. Citations in 19 academic papers and a Black Hat Europe presentation suggest the tool has become a reference point within the AI security research community.

Terms in this repo

  • MCP (Model Context Protocol) · A standard connector protocol that lets AI models plug into external tools and data
  • CVE · A publicly registered identifier for a known software vulnerability
  • Jailbreak · An attack that bypasses an AI model's safety guardrails to elicit forbidden outputs
  • Red Teaming · Proactively probing a system for weaknesses by simulating an attacker's perspective
  • F1 score · A detection accuracy metric combining precision and recall, closer to 1 is better

Repository description (English)

A full-stack AI Red Teaming platform securing AI ecosystems via Agent Scan, Skills Scan, MCP scan, AI Infra scan and LLM jailbreak evaluation.

Owner · tencent-hunyuan

Open on GitHub

Trending repos

All repos →

Latest from METAL LAB