OpenAI built a super-hacker to save us from super-hackers
openai built 'gpt-red', an adversarial llm designed to automate red-teaming and strengthen model defenses against cyberattacks like prompt injections.
openai has deployed 'gpt-red', a custom llm designed to act as a super-hacker to automate safety testing and defend against cyberattacks like prompt injections. unlike human testers, gpt-red operates in a self-play loop to rapidly discover novel attack vectors, including 'fake chain-of-thought' exploits where models are tricked by spoofed internal logs. this adversarial training process was critical to the development of gpt-5.6, which openai claims is its most secure release. the move signals a shift toward automated safety systems as ai agents gain access to sensitive computer systems, creating a risk surface too large for human teams to manage alone.