#GPT-Red

OpenAI is now using AI to attack its own AI, and it's working better than humans ever did

OpenAI's GPT-Red model, trained through self-play, successfully identifies AI vulnerabilities at 84% accuracy—far surpassing human red teamers at 13%.