AI Security

Introduction

AI introduces new attack surfaces: prompt injection, data exfiltration, jailbreaks, model abuse. Securing AI features requires defence-in-depth.

Beginner analogy: Like SQL injection but for natural language — and there's no perfect 'parameterised query' yet.

Understanding the topic

Core concepts to understand:

• Prompt injection — malicious instructions hidden in user data.
• PII redaction before sending to third-party APIs.
• Output filtering — block secrets/credentials in responses.
• Rate limiting & abuse detection.
• Tool sandboxing for agents.
• Guardrails: NeMo Guardrails, Llama Guard, Prompt Shield.

Syntax reference

Visual workflow / architecture:

Input  ─► PII redact ─► prompt-injection check ─► LLM
                                                    │
                                                    ▼
                                              Output filter
                                                    │
                                                    ▼
                                              Safe response

Real-world use

OWASP Top 10 for LLMs documents these risks. Real exploits have leaked customer data via cleverly worded prompts.

Best practices

• Never trust user input — treat it as untrusted SQL.
• Use guardrails on input + output.
• Sandbox tool execution.
• Audit log every prompt.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is prompt injection?
• Q2. How do you mitigate it?
• Q3. What is OWASP LLM Top 10?