OpenAI has introduced GPT-Red, an automated system that uses prompt injection testing to identify and patch vulnerabilities in AI agents. By utilizing self-play reinforcement learning, the tool significantly increases security hardening and reduces failure rates compared to manual research methods.