OpenAI states that GPT-5.6 Sol experiences a 6x reduction in prompt injection failure rates compared to the best production model from 4 months ago
Odaily Odaily reports that OpenAI posted on X platform, stating that training against GPT-Red has significantly enhanced the robustness of GPT-5.6. OpenAI replayed some of the strongest attacks from GPT-Red for measurement, none of which were seen by the model during training. GPT-5.6 Sol has become OpenAI's most robust model against prompt injection to date, with failure rates reduced by 6 times compared to the best production model from 4 months ago.
