AI Infra Guard

Why it's ranked here

Breadth with a backer. One Apache 2.0 repository from Tencent Zhuque Lab covers infrastructure CVE matching, MCP and skill auditing, agent testing and jailbreak evaluation, four jobs behind one interface. Detection rules are plain YAML kept in the repository rather than compiled into the binary, so rule updates ship without a rebuild and are lint-checked in CI. The skill scanner publishes its own benchmark scores, an F1 of 0.9848 with Claude Opus 4.6 on SkillTrustBench, which is a measurable claim rather than a slogan. The weak spot is deployment: the web platform has no authentication.

What's good

The standalone MCP and skill scanners write SARIF 2.1.0, the standard static-analysis format, so findings drop straight into GitHub Code Scanning, GitLab or the VS Code Problems panel. Each scanner installs on its own with pip, needs only an API key and a model name, and accepts any OpenAI-compatible endpoint, including OpenRouter. The skill scanner sorts findings into nine named risk categories, from instruction hijacking to insecure dependencies. The MCP scanner can probe a live remote server with custom headers, not only read source. Docker setup is one compose command against prebuilt images.

Tradeoffs

The README states plainly that the platform lacks an authentication mechanism and must not sit on a public network, so it belongs on a laptop or an isolated internal host. The Docker stack asks for 4GB of RAM and 10GB of disk. The language-model scanners cost tokens on every run and their accuracy depends on the model you pick: the published false-positive rate ranges from 0.0120 to 0.0723 across five models. Command-line output defaults to Chinese unless you pass the English language flag. The recommended install pipes a remote shell script straight into bash.

How to use it well

Two distinct uses. For a security team auditing what AI services are exposed internally, run the Docker stack on an isolated host and point the infrastructure scan at your network ranges. For a developer shipping agent skills or MCP servers, skip the platform and pip install the matching scanner, set the language to English, and add it to CI with SARIF output so findings appear next to code review. Pick a stronger model if false positives would block merges. It does not replace runtime guardrails: it audits code and endpoints, it does not filter live traffic.

Observed

License

Apache 2.0, with an attribution requirement for derivative works

Languages

Go core service, Python scanning agents

Interfaces

Web UI, REST and WebSocket API, Go CLI, three standalone Python CLIs

Packaging

Docker Compose with prebuilt images; aig-skill-scan on pip

Output format

SARIF 2.1.0 from the standalone MCP and skill scanners

Model support

Any OpenAI-compatible endpoint via API key and base URL

Authentication

None on the web platform; README warns against public deployment

Rule storage

YAML fingerprint and vulnerability rules in the repository