Prompt Security

Introduction

Prompt security covers injection, data leakage, output poisoning, jailbreaks and compliance. It's a discipline of its own.

Beginner analogy: Like web security — every input is an attack surface, every output is a liability.

Understanding the topic

Core concepts to understand:

• Threats: injection, jailbreak, data exfiltration, output exploitation.
• Controls: delimiters, allow-lists, output filters, sandboxing.
• Sensitive data: redact before sending; never in system prompt.
• Audit log: every prompt, response, user, time.

Syntax reference

Visual workflow / architecture:

Defence layers:
  Input  → strip/sanitise/delimit
  Prompt → fixed system, no user secrets
  Output → filter PII, validate JSON
  Action → sandbox + approval

Real-world use

Regulated industries treat prompt security as part of AppSec — same rigor as SQL injection.

Best practices

• Treat user input as untrusted always.
• Never put secrets in system prompts.
• Filter output before showing or storing.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Top prompt security threats?
• Q2. Where put secrets?
• Q3. How sandbox actions?
• Q4. Audit log fields?