Input Guardrails

Input guardrails protect your system before the model is invoked.

They validate, sanitize, and constrain incoming data to prevent unsafe or malicious inputs.

What input guardrails protect against

• Prompt injection
• Jailbreak attempts
• Excessive input size
• Secrets or credentials leakage
• PII / PHI exposure
• Unsafe or disallowed content
• Encoding and obfuscation tricks

Common input guardrails

Examples include:

• Input Size Guardrail
  Prevents extremely large inputs.

• Secrets Detection
  Blocks API keys, tokens, or credentials.

• Prompt Injection Detection
  Detects attempts to override system instructions.

• NSFW / Hate / Violence Detection
  Enforces content safety policies.

Execution behavior

Input guardrails run in this order:

1. Normalize input
2. Execute configured guardrails
3. Stop on block (if configured)
4. Emit analytics events

Example behavior

If a prompt contains an API key:

• Execution is blocked
• The request is logged
• Analytics are emitted
• A structured error is returned

Best practices

• Always include Input Size
• Always include Secrets Detection
• Use stricter policies for public APIs
• Use relaxed policies for internal tools

Next steps

• Learn about Output Guardrails
• Explore Profiles to bundle input guardrails