Note_Tech

All technological notes.

---

Project maintained by simonangel-fong Hosted on GitHub Pages — Theme by mattgraham

Guiding Principle & Guardrails

Back

• Guiding Principle \& Guardrails
  • AI Guiding Principle
  • Guardrail
    • Types of AI Guardrails
      • Input Guardrails (Pre-processing)
      • Output Guardrails (Post-processing)
      • Policy-Based Guardrails
      • Contextual Guardrails
      • Tool / Action Guardrails (Critical for AI Agents)
  • How Guardrails Are Implemented
    • Rule-Based Filters
    • ML-Based Moderation
    • Prompt Engineering
    • Human-in-the-Loop

---

AI Guiding Principle

• Veracity & Robustness
  • to achieve correct output
  • reduce hallucination
• Safety
  • prevent harmful system output and misuse
    • toxicity: the offensive, discriminatory or harmful content
• Privacy and Security
  • protect data and models
  • prevent leaks of PII data or confidential data, copyright infringement
• Explainability
  • ensure decision making can be traced and explained
• Fairness
  • ensure diversity
• Transparency
  • Communicate info about AI system to stakeholders
• Controllability
  • monitor and steer AI behavior
• Governance
  • incorporate best practices into the AI supply chain.

---

Guardrail

• Guardrails
  • rules, filters, and control mechanisms designed to ensure an AI system behaves safely, ethically, and within intended boundaries, and ensure the reliability, safety, and compliance of large language models (LLMs) by implementing validation checks on input and output data.

---

Types of AI Guardrails

ser Input → [Input Guardrails] → AI Model → [Output Guardrails] → User
                               ↓
                        [Action Guardrails]

Input Guardrails (Pre-processing)

Control what users are allowed to send into the AI.

Examples:

• Block malicious prompts (e.g., prompt injection)
• Filter sensitive data (passwords, personal info)
• Reject disallowed topics

Example Scenario:

User tries to request illegal instructions → blocked before reaching the model

---

Output Guardrails (Post-processing)

Filter or modify the AI’s response before returning it.

Examples:

• Remove harmful content
• Detect hallucinations
• Enforce tone (professional, neutral)

Example Scenario:

AI generates unsafe advice → system rewrites or blocks it

---

Policy-Based Guardrails

Rules defined by organizations or platforms.

Examples:

• No hate speech
• No illegal guidance
• Restricted domains (e.g., medical/legal advice)

Used heavily in enterprise AI systems.

---

Contextual Guardrails

Limit AI behavior based on user role or scenario.

Examples:

• Admin vs normal user permissions
• Internal tool vs public chatbot
• Region-based restrictions

---

Tool / Action Guardrails (Critical for AI Agents)

Control what actions an AI agent can take.

Examples:

• Restrict API calls
• Limit file access (read-only vs write)
• Prevent destructive operations (e.g., deleting data)

---

How Guardrails Are Implemented

Rule-Based Filters

• Regex, keyword matching
• Fast but limited

ML-Based Moderation

• Classifiers detect harmful content
• More flexible than rules

Prompt Engineering

Define behavior via system instructions.

You are a DevOps assistant.
Do not suggest destructive commands.
Do not expose secrets.

Human-in-the-Loop

Require approval for critical or high-risk actions