Guardrails and Safety Tutorials

This directory contains tutorials on implementing safety measures, content moderation, and guardrails for responsible AI deployment.

Prerequisites

Each tutorial in this directory includes its own setup and installation instructions. Please refer to the individual tutorial files for specific requirements.

Common requirements:

• Python 3.10 - 3.13
• IBM watsonx.ai account

Quick Start

1. Install dependencies (see Installation above)
2. Navigate to this directory:

   cd tutorials/07-guardrails-and-safety
   

3. Open and run your first tutorial

Tutorials

1. LLM Guardrails (llm-guardrails.ipynb)

Introduction to implementing guardrails for LLM applications.

• Topics: Input/output filtering, content moderation, safety checks
• Prerequisites: Basic LLM knowledge
• Estimated time: 30-40 minutes

2. AI Agent Security (ai-agent-security/)

Comprehensive guide to securing AI agents and multi-agent systems.

• Topics: Authentication, RBAC, secure tool execution, data protection, GDPR compliance
• Framework: IBM BeeAI
• Prerequisites: Agent development knowledge
• Estimated time: 60-90 minutes
• Files:
  • ai-agent-security-tutorial.ipynb - Main tutorial notebook
  • secure_tool_wrapper.py - Secure tool wrapper implementation
  • requirements.txt - Dependencies

3. Granite Guardian Web App (granite-guardian-webapp/)

Full-stack application demonstrating content moderation with IBM Granite Guardian.

• Topics: Real-time content filtering, web application security
• Prerequisites: Web development basics
• Type: Streamlit Application

What are Guardrails?

Safety mechanisms that:

• Validate inputs before processing
• Filter outputs before delivery
• Detect harmful content (toxicity, bias, PII)
• Enforce policies (content guidelines, compliance)
• Monitor behavior (anomaly detection, drift)

Types of Guardrails

Input Guardrails

• Prompt injection detection
• PII detection and redaction
• Content filtering
• Input validation

Output Guardrails

• Toxicity filtering
• Factuality checking
• Hallucination detection
• Bias detection
• PII redaction

Behavioral Guardrails

• Rate limiting
• Cost controls
• Topic restrictions
• Conversation monitoring

Common Use Cases

Enterprise

• Compliance (GDPR, HIPAA)
• Brand safety
• Data privacy
• Content moderation

Customer-Facing

• Child safety
• Harassment prevention
• Misinformation control
• User protection

Internal Tools

• Access control
• Audit logging
• Quality assurance
• Risk management

Additional Resources

• Main Repository README
• IBM Watsonx AI Governance
• OWASP LLM Top 10
• LangChain Safety Best Practices

Contributing

Found an issue or want to add a new guardrails tutorial? See our Contributing Guide.

License

See the LICENSE file in the repository root.