Configure prompt injection attack protection

Activate or deactivate prompt injection attack detection settings to protect all generative AI interactions on your instance from malicious inputs and unintended model behaviors.

Before you begin

Role required: sn_generative_ai.nsa_admin

About this task

AI Guardian detects and logs prompt injection attempts across all generative AI applications and features on your instance. You can configure AI Guardian to block the AI-generated response when an attack is detected.

Prompt injection detection is enabled by default for all ServiceNow Otto skills, except Platform and custom skills, which can be configured manually. The default action is block and log, with a medium severity threshold. When a skill has its own setting, AI Guardian automatically applies the more protective of the two settings, the skill-level setting or the instance-level setting.

You can export logs for review. For more information, see Export ServiceNow Otto Guardian logs.

Procedure

1. Navigate to All > ServiceNow Otto Admin > Settings.
2. In the side panel, go to ServiceNow Otto Guardian > Prompt Injection.
3. Select the Enable at instance level toggle to activate prompt injection detection at the instance, if it is not already enabled.
   
4. In the Choose an action when prompt injection is detected section, select one of the following options to handle the detected attacks:
   • To log the request and conversation while keeping the model response visible to the user, select Log the output.
   • To block the model response and log the request and conversation, select Block the response and log the output.
5. In the Select the attack severity level to check for prompt injection section, select a severity level to check for prompt injection.
   • To flag even the slightest hints of injection or manipulation attempts, select Low.
   • To flag clear or moderate prompt injection attempts, select Medium.
   • To flag only high certainty prompt injection attempts, select High.
6. In the Product-specific configuration for skills section, select the skill you want to configure.
   The Prompt injection for the selected skill page opens.
   1. In the Choose an action when prompt injection is detected section, select the detection mode for the skill:
      • To log the request and conversation while keeping the model response visible to the user, select Log the output.
      • To block the model response and log the request and conversation, select Block the response and log the output.

      Note:

      This option is only available when the instance-level action is set to Log the output.
   2. In the Select the attack severity level to check for prompt injection section, select a severity level, Low, Medium, or High.
   The skill uses the configured detection mode, or the global setting if it is stricter.
7. Select Save.

Result

Prompt injection detection is configured on your instance for all generative AI workflows. AI Guardian detects prompt injection attempts based on the severity level you selected and responds according to the action you configured. When a skill has its own setting, the more protective setting applies automatically.