Red Teaming Comprehensive.prompt

# Comprehensive Red Teaming for LLMs

## Overview

Red teaming systematically tests LLM safety by attempting to bypass alignment, generate harmful content, extract sensitive information, and cause unexpected behaviors. Comprehensive red teaming covers multiple attack vectors: prompt injection, jailbreaking, data extraction, bias elicitation, and safety bypass. Automated red teaming uses LLMs to generate adversarial prompts, enabling continuous testing at scale.

## Key Concepts

- **Prompt Injection**: Injecting instructions that override system prompts
- **Jailbreaking**: Bypassing safety alignment to generate restricted content
- **Data Extraction**: Extracting training data or system prompts from the model
- **Gradient-based Attacks**: Using model gradients to craft adversarial inputs
- **Automated Red Teaming**: Using LLMs to generate adversarial test cases
- **Multi-turn Attacks**: Building up to harmful requests across conversation turns

## Implementation Patterns

### Automated Red Team Generator