OpenAI Unveils Automated Red-Teaming AI 'GPT-Red,' Strengthening Resistance to Prompt Attacks Through Self-Play

OpenAI has unveiled “GPT-Red,” an internal-only red-teaming AI that uses self-play to automatically discover and fix model vulnerabilities. By repeating attacks that mimic the work of human red-teamers at scale and automatically, and feeding the results directly into the training process for production models, it has significantly improved resistance to prompt injection.

Details

• Training through self-play: GPT-Red is trained alongside a diverse pool of defender LLMs simultaneously. GPT-Red is rewarded when an attack succeeds, and defenders are rewarded when they successfully block one, driving both sides to improve competitively
• Automating human-like attack processes: Like a human red-teamer, GPT-Red sends prompts, observes the GPT model’s responses, and refines its next move accordingly — repeating this cycle at scale, automatically
• Direct feedback into production models: Attack patterns discovered by GPT-Red are fed directly into the training process for defensive models, continuously strengthening the robustness of models that are actually shipped
• Concrete results: The latest model, “GPT-5.6 Sol,” saw failures drop to one-sixth of the previous model’s rate on the hardest direct prompt injection benchmark, with an attack success rate as low as 0.05%
• GPT-Red itself remains private: GPT-Red is an internal-only tool, separate from any distributed model, and isn’t available for direct use by end users or developers

How to try it

GPT-Red itself is an internal, non-public tool with no plans for public release or external availability. However, its results are already reflected in models like GPT-5.6 Sol, currently available through ChatGPT and the API — you can experience the improved resistance to prompt injection in everyday use. Technical details are said to be forthcoming in a preprint paper.