Large Model Safety Workbench

Scenarios

The Large Model Safety Workbench displays the compliance status of the AI models in real time. It supports data corpus, inference services, and environment security risk operations. So, you can identify risks and potential threats to AI models in a timely manner. The Large Model Safety Workbench displays the risk status of inference security, corpus security, and environment security.

• Inference security: SecMaster analyzes WAF attack and access logs and displays the number of API calls, the number of API calls that match protection policies, the domain names of top 5 inference models matching protection policies and the number of risks, prompt injection attack distribution, and inference model attack type distribution.
• Corpus security: SecMaster analyzes DSC alarm logs and displays the corpus risk types and quantity, and the distribution of top 5 corpus risk assets.
• Environment security: SecMaster analyzes its baseline checks, vulnerabilities, and alerts and displays the top 5 compliance check risks, top 5 vulnerability risks, top 5 alerts, and recent attacks in the current workspace.

Prerequisites

• Inference security depends on the Web Application Firewall (WAF). To use Inference Security in AI Model Security module, ensure that you have a valid cloud WAF edition in use. For details, see Buying a Cloud WAF Instance.
• Corpus security depends on Data Security Center (DSC). To use Corpus Security in the AI Model Security module, ensure that the Large Model Data Security Protection you buy in DSC is still valid. For details about how to buy DSC, see Buying DSC.
• You have integrated WAF attack logs, WAF access logs, and DSC alarm logs on the SecMaster console. For details about how to access cloud service logs, see Enabling Log Access.
• Only SecMaster professional edition supports this function.

Viewing the Large Model Safety Workbench

1. Log in to the SecMaster console.
2. Click in the upper left corner of the management console and select a region or project.
3. In the navigation pane on the left, choose Workspaces > Management. In the workspace list, click the name of the target workspace.
   Figure 1 Workspace management page
   

4. In the navigation pane on the left, choose Workbenches > Large Model Safety Workbench.
5. On the Large Model Safety Workbench page, check the data on Inference Security, Corpus Security, and Environment Security.

   Table 1 Parameters on the Large Model Safety Workbench page

   Module	Parameter	Description
   Statistics Period	Statistics Period	You can select or customize a time range to check the risk overview. The options are as follows: Last 24 hours Last 3 days Last 7 days Last 30 days Custom: You can customize the start time and end time.
   Inference Security	Inference Security	The Inference Security module displays the risk operation results of the AI model inference service within the specified statistical period. Total Requests: the total number of requests for AI models in WAF request logs. Hit Protection Policies: the total number of prompt injection attack, prompt content compliance, and response content compliance alerts in WAF attack logs. Prompt Injections: the total number of prompt injection attacks identified in WAF attack logs. Prompt Content Compliance Risks: the total number of prompt content compliance alerts identified in WAF attack logs. Response Content Compliance Risks: the total number of response content compliance alerts identified in WAF attack logs.
   	Request Trends	Request Trends: displays Total Requests and Hit Protection Policies over time. The data is updated in real time. Where, Total Requests: the number of requests for AI models in WAF request logs. Hit Protection Policies: the total number of prompt injection attack, prompt content compliance, and response content compliance alerts in WAF attack logs. You can select or customize a time range to check the request trend. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   	Top 5 Assets by Risks	This chart displays the domain names of the top five inference models that hit protection policies most based on WAF attack logs. You can also learn of how many attacks or risks have been identified for these models. You can select or customize a time range to check the top 5 assets by risks. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   	Prompt Injection Distribution	This chart displays the number of prompt injection attacks by type, including jailbreak, prompt leakage, role playing, insecure command topics, insecure viewpoints, reverse inducement, inappropriate content, compliance violations, and personal data, based on WAF attack logs. You can select or customize a time range to check the prompt injection distribution. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   	LLM Attack Types	This chart displays the number of prompt injection, prompt content compliance, and response content compliance alerts identified in WAF attack logs. Attacks are displayed by type. You can select or customize a time range to check the LLM attack types. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   Corpus Security	Top 5 Corpus Assets by Risks	This chart displays the top 5 OBS buckets with the most alerts reported in DSC attack logs. The number of alerts is calculated by bucket ID. You can select or customize a time range to check the top 5 corpus assets by risks. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   	Text Risks by Category	This chart displays the number of text risks reported in DSC attack logs. Text risks include personal privacy, content compliance, and source or copyright compliance violations. You can select or customize a time range to check the text risks by category. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   Environment Security	Top 5 Compliance Checks	This chart displays the top 5 non-compliant risks with the most affected assets. You can select or customize a time range to check the top 5 compliance checks. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   	Top 5 Vulnerabilities	This chart displays the top 5 vulnerabilities within Statistical Period you specify in Large Model Safety Workbench. Vulnerabilities with the same risk severity are sorted by vulnerability name. You can select or customize a time range to check the top 5 vulnerabilities. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   	Top 5 Alerts	This chart displays the top 5 alerts by the alert discovery time. You can select or customize a time range to check the top 5 alerts. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.
   	Latest Attacks	The attack list is displayed based on the Statistics Period set on the AI Model Security page. Attacks are alarms reported by other security services configured for each defense layer in SecMaster. For details about the seven defense layers, see Overview. You can select or customize a time range to check the latest attacks. You can select Last 24 hours, Last 3 days, Last 7 days, Last 30 days, or Custom. If you select Custom, you need to specify a start and end time.