AI Security Guardrail features the Security Operations Agent, a tool built on a multi-agent architecture. The agent uses a lead agent for intelligent intent routing to manage security operations scenarios, including online testing, effectiveness evaluation, and policy tuning. This topic describes how to use the Security Operations Agent.
1. Prerequisites
Step 1: Activate the service
Activate the service on the AI Security Guardrail activation page.
Step 2: Grant permissions to a RAM user
Before using an SDK or calling an API, grant the required permissions to a RAM user. You can create an AccessKey for your Alibaba Cloud account or RAM user. An AccessKey is required to authenticate API calls. For more information, see Obtain an AccessKey.
Grant permissions to a RAM user
Log on to the RAM console using your Alibaba Cloud account.
Create a RAM user. For details, see Create a RAM user.
Grant the
AliyunYundunGreenWebFullAccesssystem policy to the RAM user. This policy grants full access to Content Moderation. For details, see Manage RAM user permissions.The RAM user can now call the Content Moderation API.
2. Billing
The Security Operations Agent is currently in public preview. During this period, the feature is free of charge. To ensure a stable experience for all users, Queries Per Second (QPS) is limited during the public preview. Plan your usage accordingly.
3. Features
The Security Operations Agent is built on an open-source framework and powered by the Alibaba Cloud Model Studio infrastructure. A lead agent recognizes user intent and schedules tasks, coordinating multiple intelligent agents to perform online testing, effectiveness evaluation, and policy tuning within AI Security Guardrail. The agent has built-in, domain-specific Skills that act as schedulable, atomic capabilities. These Skills accurately respond to product inquiries and operational requests, efficiently supporting complex security operations.
The following table describes the main features of the Security Operations Agent. These features work together to streamline security effectiveness operations, from single-query verification to comprehensive policy optimization. The list of supported security operation scenarios is continuously expanding.
Agent | Core feature | Example prompts |
Online Testing Agent (formerly Online Test) | Lets you instantly test the real-time block and pass effects of the guardrail, helping you quickly verify if protection for a single piece of content meets your expectations. |
|
Effectiveness Evaluation Agent | Lets you upload offline Excel sample files to batch-test key metrics such as block accuracy and false positive rates, and automatically generates a structured evaluation report. |
|
Policy Tuning Agent | Using a labeled sample set that you provide, the agent intelligently generates guardrail configuration settings tailored to your business scenario. You can apply these settings with a single click to quickly update and activate your guardrail policies. |
|
The Policy Tuning Agent specializes in deep policy optimization for content compliance. This provides fine-grained policy tuning capabilities to accommodate the significant differences in protection standards required by various customer scenarios. The Policy Tuning Agent does not currently support policy adjustments for other protection dimensions.
To ensure the efficient operation of the Policy Tuning Agent, we recommend using a sample size between 200 and 5,000 entries, with a nearly 1:1 ratio of black and white samples. Within these parameters, a larger number of samples typically results in better tuning.
4. Instructions
Log on to the AI Security Guardrail console.
In the left-side navigation pane, click Security Operations Agent.
You can use preset commands from the cards on the console's home page or describe your operational needs in natural language. The lead agent analyzes your intent in real time and automatically dispatches the appropriate security operations agent to answer your questions or perform tasks.
When interacting with the Security Operations Agent, you do not need to memorize fixed commands or manually switch between functions. Even within the same conversation, you can simply describe your needs in natural language. The system uses an intent-driven scheduling mechanism to automatically coordinate multiple security operations agents to seamlessly respond to various questions and tasks.
4.1 Main agents
1) Online Test Agent
The Online Testing Agent (formerly Online Test) lets you instantly test the real-time block and pass effects of the guardrail. You can select the service and test samples, and the results are displayed in a structured format after the test is complete.
2) Effectiveness Evaluation Agent
The Effectiveness Evaluation Agent helps you batch-test key effectiveness metrics for your selected service, such as block accuracy and false positive rates, and automatically generates a structured evaluation report.
To ensure evaluation efficiency and system stability, the maximum number of samples per upload is currently 5,000. The evaluation time increases with the number of samples. We recommend that you plan your sample size based on your actual needs to receive timely feedback. For larger datasets, you can submit them in batches.
You can download the sample template for evaluation within the session. The
content(samples to be detected) column is required.
3) Policy Tuning Agent
The Policy Tuning Agent intelligently generates guardrail configuration settings tailored to your business scenario. You must upload a labeled sample set. The AI learns the sample features and labeling standards from your uploaded data and then generates a configuration that fits your business needs. Click Confirm and Apply Configuration to apply the new settings with one click, quickly updating and activating your guardrail policies.
To ensure the efficiency, stability, and effectiveness of policy tuning, we recommend using a sample size between 200 and 5,000 entries. A nearly 1:1 ratio of black and white samples is also recommended. When these conditions are met, a larger and more balanced sample set typically yields better tuning results.
You can download the sample template for policy tuning within the session. The
content(samples to be detected) andlabel(labeled samples, where 1 means non-compliant and 0 means compliant) columns are required.
4.2 Agent task management
You can expand Task Management in the left panel to manage sessions.
To ensure a smooth and stable interactive experience, each UID can run a maximum of three concurrent sessions. When the concurrency limit is reached, new tasks are automatically queued and start when a session becomes available. This process prevents long waits and response delays.
4.3 Agent artifacts
Expand the task monitoring view on the right. This view shows all artifacts generated within the current session. You can preview, download, or locate an artifact's position in the conversation.