OpenAI has unveiled GPT-Red, its most advanced internal automated red-teaming model, designed to improve AI safety through self-play reinforcement learning.
GPT-Red continuously attempts to exploit vulnerabilities in defender models, particularly prompt injection attacks, while the defenders learn to resist them, creating an automated self-improvement loop for safety.
OpenAI says GPT-Red generalizes beyond its training scenarios, achieving an 84% success rate on a held-out prompt injection benchmark compared with 13% for human red-teamers. The model has already been used to strengthen GPT-5.6's defenses and is intended to help scale AI safety testing as frontier models become increasingly capable.
