Guardrails

Guardrails provide input and output validation to ensure your agents handle content safely and comply with your requirements.

Overview

The guardrail framework allows you to: - Validate inputs before sending to the LLM - Validate outputs before returning to users - Chain multiple guardrails for comprehensive protection - Use tripwires to halt execution on critical failures

Built-in Guardrails

The SDK includes 9+ production-ready guardrails:

Guardrail	Purpose	Package
PII Detection	Detect emails, phones, SSNs, credit cards	security
URL Filtering	Block/allow specific URLs	security
Secrets Detection	Prevent credential leakage	security
Profanity Detection	Filter toxic content	moderation
Moderation API	OpenAI's content moderation	moderation
Prompt Injection	Detect LLM security attacks	moderation
Content Length	Limit characters, words, or lines	content
Custom Regex	Pattern-based validation	content
Rate Limiting	Prevent abuse	ratelimit

Basic Usage

import "github.com/MitulShah1/openai-agents-go/guardrail/security"

func main() {
    // Create guardrails
    piiGuardrail := security.NewPII(security.WithTripwire(true))

    // Run agent with guardrails
    result, err := runner.Run(
        ctx,
        agent,
        messages,
        nil,
        agents.WithGuardrails([]agents.Guardrail{piiGuardrail}),
    )
}

Tripwires

Tripwires halt execution immediately on failure:

// Tripwire ON - execution stops if violated
critical := security.NewPII(security.WithTripwire(true))

// Tripwire OFF - logs violation but continues
warning := security.NewPII(security.WithTripwire(false))

Best Practices

1. Use tripwires for critical guardrails (PII, secrets, injections)
2. Log violations for non-tripwire guardrails (profanity, length)
3. Layer guardrails from most to least critical
4. Test guardrails with known violation cases
5. Monitor guardrail metrics in production