MIT Technology Review is reporting that OpenAI has developed an artificial intelligence model named GPT-Red, which functions as a "super-hacker" to bolster the defenses of its other large language models against cyberattacks. The company utilized GPT-Red as a sparring partner to train its latest flagship LLM, GPT-5.6, making it the most robust release to date.

GPT-Red automates a safety evaluation process known as red-teaming, typically performed by human testers to identify vulnerabilities and potential hijacking methods in software systems. OpenAI researchers Nikhil Kandpal and Dylan Hunn, co-creators of GPT-Red, said the tool was built to future-proof safety testing as LLMs become more complex and are used in diverse applications, including agents that interact with computer files and third-party code. They noted that GPT-Red has already uncovered new attack types.

The model primarily focuses on prompt injection attacks, where malicious instructions are hidden within text to manipulate an LLM into unintended actions, such as copying confidential data or generating harmful output. OpenAI trained GPT-Red using a self-play loop in a simulated environment, allowing it to improve its attacking capabilities while other models enhanced their defenses. Researchers said GPT-Red is highly effective and persistent in finding optimal attack methods, even discovering a novel prompt injection technique called a "fake chain of thought," which tricks an LLM into acting on spoofed information.

Jessica Ji, a senior research analyst at Georgetown University’s CSET, praised the self-play loop approach, calling the results promising. OpenAI tests showed GPT-Red was more successful at finding effective attacks than human red-teamers in a 2025 experiment. While GPT-Red significantly improved GPT-5.6's resilience compared to GPT-5, it has limitations in handling conversational and image-based attacks. OpenAI stated GPT-Red supplements human red-teamers, and the company will not release the super-hacker, confident that its extensive development makes it difficult to replicate.

Full Article: Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer