AI-Infra-Guard is a security testing platform built by Tencent's Zhuque Lab that helps teams find weaknesses in AI systems before attackers do. The idea is that as companies deploy AI models, AI agents, and related infrastructure, those systems introduce new kinds of security risks that traditional scanning tools were not built to catch. The platform bundles several scanning modes into one place. One module scans AI infrastructure components for known vulnerabilities, covering over 58 tools like model servers, agent frameworks, and databases. Another module tests AI agents and their tool integrations, called MCP servers, for security issues. A third module runs jailbreak evaluations, which means it systematically tries to get AI models to produce outputs they are supposed to refuse, then measures how often that succeeds. To run it yourself, you start it with Docker and access a web interface at port 8088. A one-line install script is also provided for convenience. The README notes the platform has no built-in authentication, so it is meant for internal or private network use only, not for public-facing deployments. The project was presented at Black Hat Europe 2025 and is actively updated, with changelog entries showing weekly or biweekly releases that expand the vulnerability database and add coverage for new AI components.
← tencent on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.