Woofun AI reports that OpenAI has introduced GPT-Red, an automated red team model designed for internal use only. The system employs self-play to generate prompt injection attacks, which are then utilized to train GPT-5.6. In unencountered scenarios, GPT-Red achieved an 84% success rate against GPT-5.1, significantly outperforming human red teams at 13%. It also compromised a vending machine agent by triggering price reductions and order cancellations, and elicited sensitive data leakage from Codex CLI. Following adversarial training, GPT-5.6 Sol reduced the attack success rate from direct GPT-Red injections to 0.05%.