OpenAI Unveils GPT-Red: Automated Red-Teaming Model for AI Security

On July 15, 2026, OpenAI announced GPT-Red, an automated red-teaming AI model designed to identify vulnerabilities in its language models, particularly against prompt injection attacks. GPT-Red achieved an 84% success rate in attack scenarios, significantly outperforming human red teamers, who managed only 13%. The model uses self-play reinforcement learning, where it attacks while other models defend, continuously improving both sides. OpenAI incorporated the findings from GPT-Red into the training of its latest model, GPT-5.6, resulting in a sixfold reduction in prompt injection failures. The system demonstrated its capabilities by manipulating an AI-powered vending machine, lowering prices and canceling orders. OpenAI has stated that GPT-Red will remain an internal tool due to its offensive capabilities. The development reflects a shift towards using AI to secure AI systems, addressing the growing complexity and attack surface of AI agents.

Key Points: • GPT-Red achieved an 84% success rate in prompt injection attacks, far exceeding human testers. • The model is designed to continuously improve AI defenses through adversarial self-play. • OpenAI will not release GPT-Red publicly due to its offensive capabilities.