OpenAI develops 'GPT-Red,' an AI that automates hacking…Strengthens model security

To bolster the cyber defense capabilities of its latest large language model, GPT-5.6, OpenAI developed and utilized a specialized hacking AI model named GPT-Red. GPT-Red was designed to automate the evaluation process of 'red teams'—expert groups that identify software vulnerabilities. This became necessary as models have become more sophisticated and are utilized in agent forms that access files and websites directly, making it difficult for humans to keep pace with every type of attack.

OpenAI designed GPT-Red to focus on prompt injection attacks, which induce abnormal behavior in models. It was trained using a 'self-play' method, where it refined its attack techniques by battling other models in environments mimicking real-world usage. Dylan Patel, the lead OpenAI researcher, revealed that GPT-Red discovered new types of attack methods that had not previously been found.

This research is expected to contribute to proactively securing security protocols in preparation for the emergence of more powerful models in the future. OpenAI anticipates that such automated security systems will serve as the technical foundation for enhancing the safety of artificial intelligence.