OpenAI Details GPT-Red Automated Red-Teaming Model

OpenAI published details of GPT-Red, an internal automated red-teaming model that attacks OpenAI's own models to find prompt injection vulnerabilities. GPT-Red beat human red-teamers 84% to 13% on prompt injection. Human red-teaming is time-intensive and does not scale, while robustness evaluations are saturated by latest models.