Before running any scan, you need to add a Target that you want to test. A target is the AI system endpoint you are assessing and can be a model, application, or agent. To connect to your target, you select a connection method that matches the platform or protocol your target uses. If your target uses a non-standard protocol or requires a custom integration, use a custom target adapter. For targets that are not publicly accessible, you can use Network Channels to reach private endpoints without opening inbound firewall ports or modifying network rules.

When you configure a target, you also provide target background information, including industry classification and use case, which AI Red Teaming uses to tailor attack scenarios to your specific environment. For target endpoints that require authentication, AI Red Teaming supports three methods: static API keys or bearer tokens in request headers, username and password credentials in the request body, and OAuth 2.0 for automatic token acquisition and refresh.

When adding a target, you also select the scan modalities that the target supports. A scan modality defines how AI Red Teaming packages and delivers attacks to the target. AI Red Teaming supports three modalities: Text (plain-text prompts), Files (adversarial content embedded in PDF or Markdown documents), and Audio (adversarial MP3 or WAV clips for voice-enabled targets). You must select at least one modality when adding a target. Enabling Files requires Text to be selected first.