OpenAI releases GPT-Red, an automated red-teaming system using self-play to improve model robustness

GPT-Red is an automated red-teaming system by OpenAI, leveraging self-play to continuously improve models through adversarial interactions. It automatically generates attack prompts to test defenses against safety, alignment, and prompt injection, then feeds weaknesses back for retraining or fine-tuning, boosting overall robustness. This shifts AI safety evaluation from primarily manual red-teaming to automated, scalable methods. For large model teams and security auditors, GPT-Red cuts the cost of repeated human penetration tests while systematically covering long-standing issues like prompt injection. However, its effectiveness depends on game environment design and attack diversity; human oversight remains essential in deployment.