OpenAI Creates GPT-Red to Enhance GPT-5.6 Security Against Prompt Injection Attacks
Covered by 2 sources
OpenAI developed GPT-Red, an AI model that autonomously identifies vulnerabilities in its systems by simulating prompt injection attacks. GPT-Red outperformed human testers in discovering new attack methods, enabling OpenAI to strengthen GPT-5.6's defenses. Due to potential risks, OpenAI has restricted access to GPT-Red while using it to scale safety alongside model advancements.