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

OpenAI disclosed GPT-Red, an automated red-teaming model that scales discovery of prompt‑injection vulnerabilities by iteratively probing LLMs to achieve malicious goals (for example, data exfiltration, credential theft, payment fraud, and remote script injection). The report describes multiple case studies where GPT-Red succeeded against earlier models, the training approach (self-play reinforcement learning vs. defender models), and claims that integration of GPT-Red into training has substantially reduced prompt-injection failures in its latest model, GPT-5.6 Sol.