Prompt Guardrails

Prompt Monitoring Guardrails set the policies that determine which activity is identified as a violation across monitored AI systems. These guardrails establish the detection rules used by Prompt Monitoring to flag prompts, responses, and conversations for review, combining standard protections with monitoring rules specific to the organisation.

The page provides controls for managing default and custom guardrails, configuring sensitivity labels, and grouping these settings into reusable Guardrail Packages. Packages can be applied across multiple AI systems, while individual systems can use targeted overrides where a different configuration is required.

Purpose

Prompt Monitoring Guardrails are used to:

• establish baseline monitoring policies across AI systems
• identify common risks through default guardrails
• define custom guardrails for organisational, regulatory, or operational requirements
• configure sensitivity labels for context-aware detection
• combine monitoring settings into reusable Guardrail Packages
• maintain consistent configurations across multiple AI systems
• apply system-specific overrides while retaining the organisation-wide baseline

The guardrails configured here determine what Prompt Monitoring detects. When monitored interactions match these policies, the resulting detections can be reviewed through the Prompt Monitoring workflow and escalated as incidents where required.

Overview of Page Sections

Guardrail Package

A monitored AI system can be linked to a Guardrail Package, which provides the baseline monitoring configuration for that system. When no package is assigned, the global default configuration applies.

Within an individual AI system, the guardrails view identifies when its configuration differs from the assigned package or the tenant-wide defaults.

Packages provide a consistent way to configure AI systems that share similar use cases, governance requirements, or risk profiles.

Default Guardrails

Default guardrails provide the standard monitoring policies available across the organisation.

They establish baseline coverage for common detection scenarios and can be applied across different monitored AI systems.

Custom Guardrails

Custom guardrails extend the default monitoring policies with rules created for specific organisational requirements, AI systems, business processes, or use cases.

They provide additional monitoring where the standard guardrails do not cover an internal policy, regulatory obligation, or specific operational requirement.

From the AI system guardrails view, users can add, edit, and remove custom guardrails.

Sensitivity Labels

Sensitivity labels are managed within the guardrails area and provide additional context for detection scenarios.

They work alongside default and custom guardrails as part of the overall monitoring configuration. Sensitivity labels can also be included within a Guardrail Package for reuse across AI systems.

Save and Reuse Configurations

A combination of default guardrails, custom guardrails, and sensitivity labels can be saved as a Guardrail Packageand reused across multiple AI systems.

From an individual AI system, users can:

• save the current configuration as a package
• update the package currently assigned to the system
• save the configuration as a new package

This provides a consistent baseline while still allowing controlled differences between individual AI systems.

Guardrail Packages

Guardrail Packages provide a scalable way to manage monitoring policies across multiple AI systems.

Instead of configuring the same policies separately for each system, selected default guardrails, custom guardrails, and sensitivity labels can be grouped into a package and applied wherever the same monitoring configuration is required.

The package management workflow allows users to:

• create Guardrail Packages
• edit existing packages
• view how many AI systems use each package

When a package is assigned, configuration changes made specifically to an AI system are treated as system-level overrides. These changes do not modify the shared package unless the user chooses to update the package itself.

When a shared package is updated, the interface identifies that the change affects other AI systems using the same package.

Prompt Policies Page

The organisation-wide management area for these configurations is available under Guardrails → Prompt Policies.

The Prompt Policies tab provides the broader view of monitoring policies across the tenant. Administrators can use this page to:

• manage global default guardrails
• create and maintain Guardrail Packages
• view the packages available across the organisation
• review which AI systems are assigned to each package
• manage custom guardrails
• manage sensitivity labels

The Prompt Policies area includes dedicated sections for Guardrail Packages and AI System Assignments, providing visibility into both the available policy configurations and where they are applied.

How This Supports Prompt Monitoring

The two areas perform separate but connected functions:

• Prompt Monitoring Guardrails determines what activity is detected
• Prompt Monitoring displays where those detections occurred and provides the workflow for reviewing them

Guardrails therefore provide the policy and detection criteria used to identify interactions that should appear within Prompt Monitoring as alerts or violations.

Notes:

• Prompt Monitoring Guardrails represents the complete monitoring policy configuration, including default guardrails, custom guardrails, sensitivity labels, and packages.
• Custom guardrails and sensitivity labels operate as components of the wider guardrail configuration.
• Guardrail Packages provide a reusable configuration for applying the same monitoring policies across multiple AI systems.
• Changes made only at the AI-system level remain overrides and do not modify the assigned package or tenant-wide defaults.
• Shared package settings change only when the package itself is explicitly updated.