OpenAI releases GPT-Red, an internal red team model for automated cyberattack simulations

OpenAI introduced its internal cybersecurity red team model, GPT-Red, on July 15. The model automates cyberattack simulations to improve the robustness of the company's products.

Every production model since GPT-5.3 has been trained using GPT-Red. The success rate of fake chain-of-thought attacks dropped from 95% on GPT-5.1 to less than 10% on the latest model. The newest GPT-5.6 Sol has only a 0.05% failure rate against direct prompt injection attacks from GPT-Red.

GPT-Red uses self-play reinforcement learning, training alongside defensive LLMs across various red team scenarios. As defensive models get stronger, GPT-Red is forced to find more powerful and diverse attacks. The red team model is isolated from product models, and OpenAI believes this starts a positive AI cybersecurity loop.