Introduction

AI Red Teaming is the practice of simulating adversarial attacks against AI systems to proactively identify vulnerabilities, potential misuse scenarios, and failure modes before malicious actors do. Distinct from traditional cybersecurity red teaming, it focuses on the unique attack surfaces of AI models, such as prompt manipulation, data poisoning, model extraction, and evasion techniques. The primary goal for an AI Red Teamer is to test the robustness, safety, alignment, and fairness of AI systems, particularly complex ones like LLMs, by adopting an attacker's mindset to uncover hidden flaws and provide actionable feedback for improvement.