// the find
Tencent/AI-Infra-Guard
A full-stack AI Red Teaming platform securing AI ecosystems via OpenClaw Security Scan, Agent Scan, Skills Scan, MCP scan, AI Infra scan and LLM jailbreak evaluation.
AI-Infra-Guard is Tencent Zhuque Lab's red teaming platform for AI systems — it scans running AI services (Ollama, vLLM, ComfyUI, etc.) for known CVEs, audits MCP servers and agent skills for supply-chain risks, and runs jailbreak evaluations against LLM endpoints. It's aimed at enterprise security teams who are deploying AI infrastructure and want automated vulnerability assessment without building the tooling themselves.
The CVE coverage is genuinely impressive — 1900+ rules across 68+ AI frameworks, with weekly updates tracked in CHANGELOG. The skill-scan module benchmarks well on SkillTrustBench (F1 ~0.98 with top models), and the taxonomy (T01–T09) gives you a structured way to reason about agent supply-chain risk. The attack library in AIG-PromptSecurity is exhaustive — 60+ encoding variants, multi-turn strategies like crescendo and tree jailbreaking — more breadth here than most open-source alternatives. Docker Compose deployment is genuinely one command.
No authentication on the web interface — the README explicitly warns not to expose it publicly, which is a significant operational constraint for any shared team deployment. The platform is deeply Tencent-ecosystem-oriented (OpenClaw, ClawHub, EdgeOne), so integrating it into a non-Tencent CI/CD pipeline requires more glue work than the docs suggest. The Python codebase has no visible test suite in the directory tree, so the 0.98 F1 benchmark numbers can't be independently reproduced without access to SkillTrustBench. Finally, the jailbreak evaluation requires you to expose your LLM's API key and endpoint in the tool's settings, which is a credential-handling risk if the instance is shared.