OpenAI details GPT-Red, an internal automated red-teaming model that helps it find and fix prompt injection vulnerabilities at scale before wider deployment (OpenAI)