GPT-Red: Unlocking Self-Improvement for Robustness

OpenAI announced on July 15, 2026, that it has trained GPT-Red, an automated red-teaming model designed to find vulnerabilities in its systems before deployment. The company said GPT-Red was trained using self-play reinforcement learning against a collection of defender models at a compute scale comparable to its largest post-training runs. OpenAI said the model can break nearly all internal and production models up to and including GPT-5.5. The company used GPT-Red to generate prompt injections for training GPT-5.6, which OpenAI said achieves six times fewer failures on its hardest direct prompt injection benchmark compared to its best production model from four months earlier. OpenAI said it keeps GPT-Red separate from deployed models to prevent malicious capabilities from reaching adversarial actors while instilling robustness in production systems. The company said it will continue to scale this approach alongside human and third-party red-teaming, layered safeguards, and real-time monitoring.