OpenAI Built an AI to Attack Itself — Meet GPT-Red, the Security Model That Breaks Things Before You Do
Reading Time: 6 minutesOpenAI has built GPT-Red, an internal AI model that automatically attacks its own systems to uncover security flaws before release, achieving an 84% attack success rate versus 13% for human red-teamers. The system focuses on prompt injection vulnerabilities and was used to train GPT-5.6 Sol, which achieved six times fewer failures on the hardest prompt-injection tests compared to OpenAI’s previous production model.