OpenAI has introduced GPT-Red, an advanced large language model (LLM) designed to act as a cyberattack simulator to improve the security of its other models. The company claims that training GPT-5.6 against GPT-Red made it its most secure release to date. GPT-Red automates red-teaming, a process typically handled by human testers, to identify vulnerabilities in software systems before they are deployed. This approach is especially important as LLMs become more complex and are used in a wider range of tasks, including interacting with computer files and third-party code. OpenAI says GPT-Red helps future-proof its safety testing process by discovering new attack methods that human teams might miss. Source: mittr

GPT-Red was developed through a self-play loop where it competed against other models to find vulnerabilities. The training environment mimicked real-world scenarios such as browsing the web, reading emails, and editing code. Over multiple rounds, GPT-Red became more effective at identifying and exploiting weaknesses, while the models it targeted improved their defenses. The system is particularly skilled at detecting prompt injection attacks, where hackers insert hidden instructions into an LLM to manipulate its behavior. OpenAI claims GPT-Red discovered a new type of prompt injection attack called a fake chain of thought, which tricks models into acting on spoofed information. Source: mittr

OpenAI tested GPT-Red against previous versions of its models, finding that more than 90% of the strongest attacks it discovered worked on GPT-5, while fewer than 23% were effective on the newer GPT-5.6. Despite its capabilities, GPT-Red is not perfect and struggles with certain types of attacks, such as those involving back-and-forth conversations between attackers and targets. It also has limitations in handling image-based attacks, which can be used to pass text in prompt injection scenarios. OpenAI says GPT-Red complements human red-teamers, allowing them to focus on areas where their expertise is most needed. Source: mittr