OpenAI introduced GPT-Red, an internal model designed to automate red-teaming security assessments. The system identifies security flaws in other AI models by simulating malicious attacks.
GPT-Red specializes in detecting prompt injection vulnerabilities. These flaws occur when malicious instructions hidden in third-party data manipulate AI behavior.
The model generates high volumes of attack scenarios to improve testing efficiency over manual human methods. OpenAI uses this data to adversarially train upcoming systems like GPT-5.6 before public release.