What AI Models Does AI-Infra-Guard Support? A Complete Guide to LLM Integration

AI-Infra-Guard supports major commercial LLMs including OpenAI GPT-4, Anthropic Claude, Google Gemini, Alibaba/Kimi, DeepSeek, and Zhipu AI GLM, plus any self-hosted model with an OpenAI-compatible REST API endpoint.

Tencent's AI-Infra-Guard (A.I.G) is an open-source AI infrastructure security scanning platform designed to integrate seamlessly with diverse large language models. The architecture treats any model accessible via a standard API endpoint as a first-class citizen, allowing security teams to leverage both cloud-based providers and private deployments for vulnerability detection tasks.

Commercial LLM Providers Supported by AI-Infra-Guard

The platform ships with pre-configured support for the industry's leading LLM families. These integrations are defined in the source code and validated through the aig-skill-scan benchmark suite.

OpenAI Models

According to common/utils/models/openai.go, AI-Infra-Guard natively supports OpenAI's model family including gpt-4 and gpt-3.5-turbo. The implementation uses a standardized struct that expects model, token, and base_url parameters, making it compatible with both OpenAI's official API and compatible proxies.

Anthropic Claude

The README's performance benchmark table lists claude-opus-4.6 as a tested and validated model, achieving an F1 score of 0.9848 with high precision (0.9725) and recall (0.9974) in security scanning tasks.

Google Gemini

gemini-3.5-flash is officially supported and benchmarked, showing the lowest false positive rate (0.0120) among tested models while maintaining strong precision at 0.9947.

Alibaba Cloud and Moonshot AI (Kimi)

The platform includes dedicated support for kimi-2.6 (Moonshot AI) and related Alibaba Cloud models, achieving an F1 score of 0.9780 in security scanning evaluations.

DeepSeek

deepseek-v4-flash is fully supported with demonstrated performance metrics showing an F1 score of 0.9740, making it suitable for high-throughput security analysis workflows.

Zhipu AI (GLM)

glm-5.1 achieves performance parity with top-tier models, scoring 0.9836 F1 with particularly strong recall metrics (0.9974) for detecting AI infrastructure vulnerabilities.

Self-Hosted and Custom AI Model Support

Beyond commercial providers, AI-Infra-Guard supports any self-hosted or custom model exposing an OpenAI-compatible REST API. This includes deployments via Ollama, vLLM, NVIDIA Triton, ComfyUI, and n8n.

The generic model schema lives in pkg/database/model.go, which defines the storage structure for user-defined configurations. The system loads these definitions dynamically from db/model.yaml (handled in pkg/database/yaml_model.go), allowing runtime integration of private models without code changes.

How AI-Infra-Guard Configures AI Models

The platform uses a unified three-field schema to standardize LLM integration across all providers. As implemented in common/utils/models/openai.go, every model configuration requires:

{
  "model": "your-model-name",
  "token": "your-api-key-or-token",
  "base_url": "https://your-endpoint/v1"
}

The ModelStore implementation in pkg/database/model.go persists these entries using GORM, storing ID, name, token, base URL, and rate limits. At runtime, common/websocket/task_manager.go injects these parameters into task execution contexts, enabling dynamic model switching per scan.

Model Performance Benchmarks

The following table from the repository README shows validated AI models and their performance characteristics in the aig-skill-scan component:

| # | Model | F1 | Precision | Recall | FPR |

|---|--------------------------|--------|-----------|--------|-----| | 1 | Claude Opus 4.6 | 0.9848 | 0.9725 | 0.9974 | 0.0663 | | 2 | GLM 5.1 | 0.9836 | 0.9701 | 0.9974 | 0.0723 | | 3 | Gemini 3.5 Flash | 0.9792 | 0.9947| 0.9641 | 0.0120 | | 4 | Kimi 2.6 | 0.9780 | 0.9895 | 0.9667 | 0.0241 | | 5 | DeepSeek v4 Flash | 0.9740 | 0.9868 | 0.9615 | 0.0301 |

These benchmarks demonstrate that AI-Infra-Guard supports models ranging from 0.9740 to 0.9848 F1 scores, ensuring high-quality security analysis regardless of the chosen provider.

Running Infrastructure Scans with Specific AI Models

Command-Line Interface

Execute a skill scan using a specific supported model via the CLI:

aig-skill-scan --repo ./my-skill \
                -m deepseek-v4-flash \
                --language en \
                -o result.json

HTTP API Integration

Submit scanning tasks programmatically using the JSON payload format processed by common/websocket/task_manager.go:

{
  "type": "ai_infra_scan",
  "content": {
    "target": ["http://127.0.0.1:8000"],
    "model": {
      "model": "gpt-4",
      "token": "sk-xxxxxxxx",
      "base_url": "https://api.openai.com/v1"
    }
  }
}

This payload structure applies universally across all supported AI models, whether commercial or self-hosted.

Summary

• AI-Infra-Guard supports six major commercial LLM families: OpenAI, Anthropic, Google, Alibaba/Kimi, DeepSeek, and Zhipu AI (GLM).
• The generic API schema in pkg/database/model.go enables integration with any OpenAI-compatible endpoint, including Ollama, vLLM, and custom deployments.
• Model configurations require only three parameters: model, token, and base_url, stored persistently in db/model.yaml.
• Benchmarked models achieve F1 scores between 0.9740 and 0.9848, with Claude Opus 4.6 and GLM 5.1 showing the highest accuracy.
• The task manager (common/websocket/task_manager.go) dynamically injects model parameters into scanning workflows via CLI or HTTP API.

Frequently Asked Questions

Can I use AI-Infra-Guard with a local LLM running on my own server?

Yes. AI-Infra-Guard supports any model exposing an OpenAI-compatible REST API. Configure your self-hosted model (such as those running on Ollama, vLLM, or Triton) by specifying its base_url in the model configuration JSON. The system loads these definitions from db/model.yaml at runtime.

Does AI-Infra-Guard require separate API implementations for each LLM provider?

No. The platform uses a unified model interface defined in common/utils/models/openai.go. All providers use the same three-field schema (model, token, base_url), meaning you can switch between OpenAI, Anthropic, or custom models without changing your integration code.

Which AI model provides the best accuracy for AI-Infra-Guard security scans?

According to the benchmark data in the README, Claude Opus 4.6 achieves the highest F1 score (0.9848) and perfect recall (0.9974), making it the most accurate for detecting AI infrastructure vulnerabilities. However, Gemini 3.5 Flash offers the lowest false positive rate (0.0120) if precision is your primary concern.

How do I add a new custom model to AI-Infra-Guard?

Add your model to the ModelStore by creating an entry in db/model.yaml or via the database interface in pkg/database/model.go. Specify the model name, API token, and base_url. The system will automatically make the model available for use with aig-skill-scan commands and HTTP API tasks without requiring a restart.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}