AI-Infra-Guard

AI-Infra-Guard is Tencent's open-source security platform for testing AI applications, MCP servers, agent skills, model integrations, and the surrounding infrastructure that now sits around production AI systems. The project describes itself as a full-stack AI Red Teaming platform with Agent Scan, Skills Scan, MCP scan, AI Infra scan, and LLM jailbreak evaluation. That makes it useful for teams that need a practical scanner before connecting new agents, skills, or MCP servers to internal systems. The platform is aimed at security engineers, AI platform teams, and developers who have moved beyond prompt experiments and now need to check the attack surface created by tool use. Its README and documentation point to several workflows: scanning remote MCP server repositories or uploaded source archives, evaluating model jailbreak risk, running agent and skill scans, using API documentation after deployment, and extending checks through fingerprint, vulnerability, MCP, and evaluation datasets. AI-Infra-Guard is not just a library. The documented install paths include Docker Compose using prebuilt images, a one-line Docker startup script, source-based Docker Compose, a clawhub install command, and a separate aig-skill-scan Python package for focused skill scanning. That gives teams several ways to try the system: run the full web platform, call APIs behind an internal scanner, or add a narrower skill scan step to review Claude-style skills before they reach users. The strongest reason to evaluate AI-Infra-Guard is coverage. The project explicitly names prompt injection, jailbreak evaluation, MCP scan, skill scanning, AI infrastructure scanning, agent security, and vulnerability rules. Those are the exact areas where classic web scanners miss AI-specific risk. Tencent's repository also showed more than 5,300 GitHub stars, 500 forks, Apache-2.0 licensing, recent commits, and public docs at the time of creation, which suggests active maintenance and broad community interest. The main caveat is operational weight. A platform that scans agents, MCP servers, skills, model endpoints, and infrastructure needs careful configuration. Teams should decide whether they want the full server, the CLI scanner, or a narrower CI step before rolling it out. Pricing is free for the open-source software itself; costs come from the infrastructure used to run scans and any target LLM APIs configured for jailbreak evaluation. A sensible rollout starts with a non-production repository or archived MCP server, then expands to internal agent skills and model endpoints after the team understands the output format. Use the documentation and API pages to decide which scans should run manually, which belong in CI, and which require security-team review before developers can accept the findings.