OpenAI's GPT-Red beats human red-teamers 84% to 13% on prompt injection

GPT-Red, an internal automated red-teaming model from OpenAI, achieved an 84% success rate on prompt injection attacks compared to human red-teamers' 13%. The model is used internally to automatically find vulnerabilities in OpenAI's own models. OpenAI published details to emphasize the need for scalable automated safety testing.