OpenAI has unveiled GPT-Red, an automated red-teaming model. It learns to design prompt injection attacks through self-play, then uses the vulnerabilities it finds to train GPT-5.6. GPT-Red is for internal use only and won’t be released publicly.
In a set of new scenarios not included in training, GPT-Red achieved an 84% success rate attacking GPT-5.1, compared to just 13% for human red teams. It also broke into a vending machine agent, making it lower prices and cancel other people’s orders. When testing Codex CLI, it was able to trick the agent into exfiltrating sensitive data.
After adversarial training, GPT-5.6 Sol’s attack success rate dropped to 0.05% when facing direct prompt injection from GPT-Red.