guardrails-safety_skill

Overview

This skill protects AI applications with layered guardrails for inputs and outputs, focusing on toxicity, PII, prompt injection, and constitutional AI policies. It provides detection, sanitization, and revision mechanisms to reduce misuse and harmful behavior. The goal is to keep models compliant, safe, and auditable while preserving useful functionality.

How this skill works

Input guardrails run classifiers and pattern detectors to flag toxicity, find and redact PII, and detect injection attempts before the model sees user content. Output guardrails evaluate responses for toxicity, factuality against context, and citation validity, and can block or flag problematic outputs. A constitutional filter applies safety principles via critique-and-revise loops to rewrite responses that violate defined rules.

When to use it

• Before sending user text to a model to prevent harmful prompts or leaked sensitive data
• When returning model responses to users to ensure no toxic or fabricated content is delivered
• During deployment of high-risk features such as code generation, medical or legal advice
• When compliance and auditability are required (PII protection, logging of blocked events)
• While running adversarial testing or red-team scenarios to validate defenses

Best practices

• Apply defense in depth: combine toxicity models, regex detectors, and constitutional review
• Log all blocked and sanitized events for auditing and continuous improvement
• Fail closed: block or quarantine outputs when safety checks are uncertain
• Continuously update and expand detection patterns and thresholds based on adversarial testing
• Run regular adversarial and regression tests to validate changes to guardrails

Example use cases

• Sanitizing chat inputs to remove emails, phone numbers, and credit card numbers before processing
• Blocking or rewriting prompts that try to bypass instructions using injection patterns
• Preventing a model from giving harmful instructions by applying constitutional principles and revising unsafe responses
• Checking generated content against source documents to flag potential hallucinations or bad citations
• Instrumenting customer-facing assistants with safety logging and automatic redaction for compliance