Guardrails Overview
Guardrails are runtime enforcement units that protect your AI system from unsafe, insecure, or non-compliant behavior.
They execute before, during, and after model execution to ensure inputs, outputs, and tool usage follow defined rules.
What is a guardrail?
A guardrail is a self-contained policy check that can:
- Block execution
- Modify content (redaction, sanitization)
- Emit warnings
- Record telemetry
- Trigger alerts or actions
Guardrails are deterministic, auditable, and configurable.
Where guardrails run
Guardrails are applied at three primary stages:
-
Input guardrails
Validate and sanitize user or system input before model execution. -
Output guardrails
Inspect and enforce policies on model responses. -
Tool guardrails
Control and restrict tool or agent behavior.
Key characteristics
Modular
Each guardrail is independent and reusable across profiles.
Composable
Multiple guardrails can be combined into profiles.
Configurable
Each guardrail accepts runtime configuration.
Observable
Every execution produces analytics and audit events.
Guardrails vs filters
| Guardrails | Filters | | ------------------- | ------------- | | Runtime enforcement | Static rules | | Context-aware | Pattern-based | | Auditable | Opaque | | Configurable | Hardcoded |
Next steps
- Learn how Input Guardrails work
- Explore Output Guardrails
- Understand Tool Guardrails
- Learn how to write Custom Guardrails