OpenAI has developed and deployed an internal artificial intelligence system, named GPTRed, that autonomously attacks its own models to discover and patch security vulnerabilities before they can be exploited. The company detailed the system this week, revealing it was instrumental in training its latest GPT-5.