OpenAI’s GPT-Red Automates Prompt Injection Testing to Harden GPT-5.6 Sol

OpenAI has developed an internal automated red-teaming model called GPT-Red to identify prompt injection vulnerabilities. The goal of GPT-Red is to discover and fix issues before the tools are widely deployed. GPT-Red is used to adversarially train OpenAI's models, with a focus on hardening GPT-5.6 Sol. OpenAI found that its previous models were highly vulnerable to GPT-Red's prompt injection attacks, indicating the effectiveness of the red-teaming model.