OpenAI Launches AI Safety Flywheel: How GPT-Red Redefines Model Robustness

To counter security weaknesses from AI's deep integration with browsers and files, OpenAI introduced GPT-Red, an automated red team model. Using self-play, it lowers direct prompt injection attack failure rate to 0.05%, outperforming manual testing in efficiency and coverage. This paves a new way for AI self-improvement and security strengthening.....