Check AI agent and infrastructure security gaps with a red team tool
When an AI agent calls external tools or a model produces unexpected responses, you can directly investigate the security holes behind them. This tool scans five areas—AI infrastructure, agents, models, and related tools—to visualize potential risks. By performing complex security tests on a single platform, it helps teams operating or developing AI systems identify risk factors in advance.
Integrated checks from five scanners
This platform performs security checks across various layers of the AI ecosystem. The AI infrastructure scan analyzes domains or IP addresses to match fingerprints and vulnerabilities. The LLM security assessment tests jailbreak scenarios on models exposing OpenAI-compatible APIs. The MCP scan checks the source directory or running services of MCP servers, while the skill scan inspects local agent skill directories. Finally, the agent scan sends test prompts to running agents to check if they trigger dangerous behaviors.
Platform deployment and individual tool usage
Users can use the service integratively through a web interface or run each scanner independently. To use the web interface, you must deploy the service via Docker, requiring Docker 20.10 or higher, at least 4GB of RAM, and at least 10GB of disk space. After deployment, you can access the web interface on port 8088 of localhost. To use it as an individual tool, install each scanner package with pip and execute the command by specifying the target address. For example, the infrastructure scan is performed with the aig-infra-scan command, and the model assessment with the aig-model-scan command.
Plugin-based extension structure
One of the core strengths of this tool is its plugin system. The community can add new fingerprint rules, vulnerability scan rules, MCP security rules, and model evaluation datasets in the form of YAML files. These rules are stored in the data/fingerprints/, data/vuln/, data/mcp/, and data/eval directories; you create files by referring to the existing format and submit them via a pull request. This allows new vulnerabilities or evaluation scenarios to be reflected quickly, continuously expanding the tool's coverage.
Items to verify before internal use
This project is an AI red team platform intended for internal use by companies or individuals. Since there is currently no authentication mechanism, it should not be deployed on public networks. It is provided as open source under the Apache License 2.0, and if you integrate the code or perform derivative work, you must retain the original LICENSE and NOTICE files. Additionally, you must state in product documentation or UI that 'Tencent Zhuque Lab has integrated AI-Infra-Guard, which was released as open source.' When citing in research or reports, you are also required to disclose the source along with the repository link.
