[OpenAI Trains AI to Attack Its Own Models, Achieves 84% Success Rate, Far Surpassing Humans] According to monitoring by Dongcha Beating, OpenAI has unveiled an automated red team model called GPT-Red. It learns to design prompt injection attacks through self-play and uses the discovered vulnerabilities to train GPT-5.6. GPT-Red is for internal use only and will not be made publicly available. In a set of new scenarios not included in training, GPT-Red achieved an 84% success rate in attacking GPT-5.1, compared to just 13% for human red teams. It also managed to compromise a vending machine Agent, causing product price reductions and the cancellation of others' orders. When testing Codex CLI, it was able to induce the Agent to leak sensitive data. After adversarial training, GPT-5.6 Sol reduced the success rate of direct prompt injection attacks by GPT-Red to 0.05%. [Original Link]