Understand Agent Hub Guardrails
Guardrails help protect Agent Hub interactions by detecting or preventing specified types of unwanted content or behavior. Agent Hub administrators configure guardrails that apply to the Agent Hub orchestrator.
Agent Hub provides guardrails for personally identifiable information (PII), content moderation, and prompt injection detection.
Changes to guardrail settings take effect immediately after you save them.
Personally Identifiable Information
The Personally identifiable information (PII) guardrail helps protect sensitive information by detecting PII in user input and generated responses. PII can include information such as names, email addresses, telephone numbers, and physical addresses.
When protection is enabled, the guardrail helps prevent sensitive information from being exposed during Agent Hub interactions.
Content Moderation
The Content moderation guardrail detects potentially harmful or inappropriate content in user input and generated responses. This can include hateful, sexual, violent, toxic, derogatory, or harassment-based content.
When protection is enabled, the guardrail helps prevent content that meets the detection criteria from being processed or generated.
Prompt Injection Detection
The Prompt injection detection guardrail detects potentially malicious instructions in user input that attempt to override an AI system's intended instructions or behavior.
When protection is enabled, the guardrail helps prevent detected prompt injection attempts from affecting request processing.
Guardrails and Other Platform Protections
Guardrails configured in Agent Hub Administration apply to the Agent Hub orchestrator. They don't replace or control the guardrails configured for individual agents.
When Agent Hub invokes an agent, that agent's own guardrail configuration applies to its processing. The protections provided by an agent depend on how that agent is configured.
Why this matters
Guardrails provide administrators with controls for protecting Agent Hub interactions from specific types of unwanted content or behavior. Understanding what each guardrail protects against can help administrators determine which protections are appropriate for their organization's Agent Hub environment.
For information about changing guardrail settings, see Configure Agent Hub Guardrails.