AIAny
AI Infra2024
Icon for item

AI-Infra-Guard (A.I.G)

Full-stack AI red‑teaming platform that fingerprints AI infrastructure for known CVEs, audits MCP servers and agent skills with LLM-driven analysis, and runs cross-model jailbreak evaluations; designed for hands-on security assessment of AI deployments.

Introduction

Why this matters

AI deployment stacks expose heterogeneous risks—from vulnerable servlets and inference endpoints to malicious agent skills and untrusted MCP connectors. This project treats that surface as layered and matches detection paradigms to each layer: deterministic fingerprinting for infra CVEs, agentic LLM analysis for semantic code/skill threats, and large-scale jailbreak enumeration for model alignment robustness.

What Sets It Apart
  • Layer-matching approach: uses deterministic fingerprint rules for components and CVEs, LLM-driven agentic audits for MCP/skill semantics, and a jailbreak harness for alignment testing — so findings are mapped to the kind of evidence that best supports remediation.
  • Broad coverage and telemetry: fingerprints for 100+ AI components and a vulnerability database covering thousands of CVE rules (project has reported >1600–2000+ rules across releases), plus specialized scanners for Ollama, ComfyUI, vLLM and others, enabling automated matching of live service fingerprints to known flaws.
  • Agent- and skill-focused tooling: standalone CLIs and agent-scan/skill-scan modules that evaluate tool-poisoning, code vulnerabilities, credential exfiltration and other agentic threats, with integration points for CI/CD and skill marketplaces.
  • Practical operator UX: Docker-based deployment, web UI at localhost:8088, API checker and plugin-driven rule updates to let teams adapt fingerprints and datasets without core changes.
Who it's for — and tradeoffs

Great fit if you operate or evaluate LLM/agent deployments and need reproducible, layered security tests (infrastructure CVE checks, MCP/skill audits, and jailbreak benchmarking) that can be automated in CI or run interactively. It is production-focused for internal networks and research workflows.

Look elsewhere if you need a turnkey SaaS with built-in authentication for public deployment — the project notes it lacks a production authentication layer and is intended for internal/private use unless hardened and wrapped by an operator. Also expect engineering effort to tune rule plugins and LLM backends for skill/MCP semantic scans.

Information

  • Websitegithub.com
  • OrganizationsTencent Zhuque Lab, Tencent Security Platform Department, Tencent
  • AuthorsYong Yang, Xing Zheng, Nicky, Zona, Fyoung, Xiangfan, Elwood, Robert, Zoe
  • Published date2024/12/25

More Items

Evaluates how well models and coding agents optimize operators for real Ascend NPU workloads, accounting for both speedups and token cost. Test cases preserve tensor shapes and data types from real model execution contexts.

Details the Xing 4 model architecture, Ascend-scale pretraining systems, and agentic post-training, including a 285B-parameter training run on 8,192 Ascend 910C cards.

Hugging Face
AI Model2026

Combines scientific data understanding with search, coding, and tool use for multi-step research. It checks intermediate results and visual feedback to revise its actions, and accepts more than ten scientific data types.