OpenAI says training against GPT-Red made GPT-5.6 about 6× more resilient to prompt injections than its best model from four months earlier

OpenAI said on Tuesday that training against GPT-Red made GPT-5.6 about six times more resilient to prompt injections than its best production model from four months earlier. The company posted on X that it measured the improvement by replaying some of GPT-Red's strongest attacks, none of which the models had seen during training. OpenAI said GPT-5.6 Sol proved to be its most resilient model against prompt injections to date, with six times fewer failures than the previous best model. The announcement did not specify the exact date of the comparison model or provide additional technical details about the training methodology. OpenAI did not immediately respond to requests for further comment.