# Guardrails

> A defensive pattern where inputs and outputs are inspected by dedicated safety agents or rules to preventing malicious use, jailbreaks, and harmful content. Use when user asks to "add safety checks", "set up guardrails", "prevent harmful outputs", or mentions agent boundaries, output validation, or content filtering.

- Skill: `lauraflorentin/guardrails` (Agent Skill)
- Install (CLI): `npx skillmds@latest add lauraflorentin/guardrails`
- Raw SKILL.md: https://api.skillmd.com/api/skills/lauraflorentin/guardrails/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: lauraflorentin (https://skillmd.com/u/lauraflorentin)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/lauraflorentin/guardrails

---


# Guardrails & Safety

Guardrails are the firewall of an AI system. They sit between the user and the agent (Input Guardrail) and between the agent and the user (Output Guardrail). They enforce policy, security, and tone. Unlike the main agent, which tries to be helpful, the guardrail tries to be safe and compliant.

## When to Use

-   **Jailbreak Prevention**: Stopping users from tricking the model ("Ignore previous instructions...").
-   **PII Protection**: Detecting and redacting phone numbers, emails, or credit cards.
-   **Topic Adherence**: Ensuring a customer support bot doesn't discuss politics or religion.
-   **Brand Safety**: preventing the model from generating offensive or competitor-promoting content.

## Use Cases

-   **Input Filter**: Blocking prompts that violate usage policies.
-   **Output Filter**: Blocking model responses that contain hate speech or hallucinations.
-   **Sandboxing**: Ensuring code generated by the agent acts within safe bounds (e.g., no network access).

## Implementation Pattern

```python
def guarded_execution(user_input):
    # Layer 1: Input Guardrail
    # Check for prompt injection or policy violations
    if not safety_agent.check_input(user_input).safe:
        return "I cannot answer that request."
        
    # Layer 2: Main Execution
    response = main_agent.run(user_input)
    
    # Layer 3: Output Guardrail
    # Check for PII or harmful content in the response
    if not safety_agent.check_output(response).safe:
        log_violation(user_input, response)
        return "Response withheld due to safety policy."
        
    return response
```


## Troubleshooting

| Problem | Cause | Fix |
|---|---|---|
| Guardrail blocks legitimate requests | Over-broad pattern matching | Tune guardrail thresholds using a labeled test set; track false positive rate |
| Agent bypasses guardrails | Prompt injection in user input | Apply guardrails *before* injecting user content into agent context |
| Guardrail adds too much latency | Synchronous pre-call check | Run guardrail in parallel with the first LLM call; cancel if flagged |
| Silent failures | Guardrail raises exception but agent continues | Treat guardrail exceptions as hard stops; log and escalate |