AI Guardrails:API integration guide

Prerequisites

Before you begin, decide the following:

• Which content to check: user inputs (query_security_check_intl), LLM outputs (response_security_check_intl), or both

• How to handle each risk level: block high-risk content automatically, route medium-risk content to human review, and treat low-risk content as safe unless you need high recall

Then make sure you have:

• An Alibaba Cloud account with the Guardrails service activated

• A Resource Access Management (RAM) user with the AliyunYundunGreenWebFullAccess system policy and an AccessKey pair (see Set up a RAM user)

Set up a RAM user

The AccessKey pair is used for identity verification when calling Alibaba Cloud API operations.

1. Log on to the RAM console with your Alibaba Cloud account.

2. Create a RAM user. For details, see Create a RAM user.

3. Grant the AliyunYundunGreenWebFullAccess system policy to the RAM user. For details, see Grant permissions to a RAM user.

4. Create an AccessKey pair for the RAM user. For details, see Obtain an AccessKey pair.

Install the SDK

For SDK installation and setup, see the SDK Reference.

API reference

Handle moderation results

Use the top-level fields (RiskLevel, SensitiveLevel, AttackLevel) to route content. Drill into Result, SensitiveResult, and AttackResult arrays for the specific labels and confidence scores that explain the decision.

A custom dictionary match always sets RiskLevel to high.

Example

Request

{
    "Service": "query_security_check",
    "ServiceParameters": {
        "content": "testing content",
        "chatId":"ABC123"
    }
}

Response (system policy matched)

{
    "Code": 200,
    "Data": {
        "Result": [
            {
                "Label": "political_entity",
                "Description":"Suspected political entity",
                "Confidence": 100.0,
                "RiskWords": "Word A,Word B,Word C"
            },
            {
                "Label": "political_figure",
                "Description":"Suspected political figure",
                "Confidence": 100.0,
                "RiskWords": "Word A,Word B,Word C"
            }
            {
                "Label": "customized",
                "Description": "Hit custom dictionary",
                "Confidence": 100.0,
                "CustomizedHit": [
                     {
                        "LibName": "Custom Dictionary Name 1",
                        "KeyWords": "Custom Keyword"
                     }
                ]
             }
        ],
         "SensitiveResult": [
            {
                "Label": "1780",
                "SensitiveLevel": "S4",
                "Description":"Credit card number",
                "SensitiveData": ["6201112223455"]
            }
        ],     
         "AttackResult": [
            {
                "Label": "Indirect Prompt Injection",
                "AttackLevel": "high", 
                "Description":"Indirect prompt injection",
                "Confidence": 100.0
            }
        ],   
        "RiskLevel": "high",
        "SensitiveLevel": "S3",
        "AttackLevel": "high",                      
    },
    "Message": "OK",
    "RequestId": "AAAAAA-BBBB-CCCCC-DDDD-EEEEEEEE****"
}

Status codes

What's next

• Multimodal API integration guide — extend moderation to AI-generated images and files

• SDK Reference — SDK installation and usage

• Billing overview — understand how requests are billed