OpenAI Builds AI That Attacks Its Own Models

OpenAI has developed GPT-Red, an internal AI system designed to attack its own models and expose prompt injection vulnerabilities before they reach users. Using self-play reinforcement learning, GPT-Red acts as an attacker against defender models, succeeding on 84% of scenarios compared to 13% for human red-teamers. The tool has already shown results, reducing a class of attacks that worked 95% of the time against GPT-5.1 to under 10% against GPT-5.6. OpenAI will keep GPT-Red internal and will not release it publicly, feeding its findings back into model training instead.