OpenAI's GPT-Red Model Attacks GPT-5.1 with 84% Success Rate, Reduces Adversarial Vulnerability to 0.05%

According to monitoring by Beating, OpenAI has unveiled GPT-Red, an automated red-teaming model designed to identify and exploit security vulnerabilities in its AI systems. GPT-Red learns to craft prompt injection attacks through self-play, with findings fed into training GPT-5.6.

In new, unseen scenarios, GPT-Red achieved an 84% attack success rate against GPT-5.1, compared to just 13% for human red teams. After adversarial training, GPT-5.6 Sol reduced the success rate of direct prompt injection attacks to 0.05%. GPT-Red remains for internal use only.