GPT-Red: an LLM super-hacker OpenAI built to make its models safer (www.technologyreview.com)

OpenAI has introduced GPT-Red, a specialized LLM designed to enhance the security of its models through automated red-teaming—a process traditionally reliant on human testers to uncover vulnerabilities. By training GPT-Red in a self-play loop with its models, it has emerged as a highly effective attacker, identifying new types of prompt injection attacks, such as a novel "fake chain of thought" manipulation. This innovative approach significantly improves the robustness of OpenAI's latest LLM, GPT-5.6, making it the most secure version to date. The significance of GPT-Red lies in its ability to adaptively discover vulnerabilities that human testers might overlook, a critical feature as LLMs become increasingly complex and integrated into more applications. During tests, GPT-Red outperformed previous human efforts in detecting weaknesses, revealing that more than 90% of its attacks could exploit flaws in the earlier GPT-5 model, while under 23% succeeded against GPT-5.6. However, while GPT-Red represents a major advancement in AI security, OpenAI acknowledges that human expertise remains vital for certain nuanced attack scenarios, ensuring a balanced approach to safeguarding AI technologies.