OpenAI details GPT-Red, an AI system designed to find vulnerabilities in its own models

GPT-Red functions as an automated red team, employing self-play reinforcement learning to continuously attack target models.