Course contentsShow
← AI Engineering
Lesson 1485 of 1,88636. Security in AI SystemsFree lesson

Input Sanitization Strategies

Implement pre-processing pipelines to sanitize user inputs before sending them to LLMs.

Input Sanitization Strategies

What you'll learn: How to build pre-processing pipelines that clean and validate user inputs before they reach your LLM, preventing injection attacks and data leakage.

The Core Problem

When users send text to your AI system, you can't trust it blindly. Malicious inputs might contain hidden instructions, attempt to override system prompts, or embed sensitive data extraction commands. Input sanitization is your first line of defense—think of it as a security checkpoint that inspects and cleans every piece of user text before it reaches your LLM.

Pre-Processing Pipeline Architecture

A robust sanitization strategy involves multiple layers of filtering:

1. Format Normalization

Strip or encode potentially dangerous characters: excessive whitespace, special Unicode sequences, control characters, and null bytes. Convert everything to a standard encoding (UTF-8).

2. Structural Validation

Check that inputs conform to expected patterns—length limits, allowed character sets, required fields. Reject malformed requests early.

3. Content Extraction

If users can submit formatted text (markdown, HTML), parse it safely and extract only the plain text content. Never pass raw markup directly to your LLM.

4. Instruction Marker Removal

Strip common injection patterns like repeated delimiters (---, ###), role-switching phrases ([SYSTEM], Assistant:), or suspicious instruction keywords. This builds on your knowledge of prompt injection vectors.

5. Semantic Pre-Validation

Use a lightweight classifier to detect obvious attack patterns before expensive LLM processing. Flag inputs that look like jailbreak attempts or system prompt extraction.

Implementation Pattern

Your sanitization pipeline should be a separate module that runs before prompt construction, logging what was modified for audit purposes. Combine regex-based filtering with NER models (like you used for PII detection) to catch sophisticated attacks.

Key Takeaway: Input sanitization is a multi-layered defense that normalizes, validates, and filters user text before LLM processing—reducing attack surface and preventing most injection attempts at the gate.