Meet GPT-Red, the LLM super-hacker that OpenAI built to future-proof its models against cyberattacks. By training its flagship model, GPT-5.6, against GPT-Red, they've made a robust release indeed.

GPT-Red automates red-teaming, typically done by humans, to find system vulnerabilities before the software is released. As LLMs get more complex and interactive, this self-play loop ensures OpenAI stays ahead of potential threats.

The researchers say GPT-Red has already found new attack vectors, like a fake chain of thought that tricks models into acting on false information. It’s particularly good at identifying the most efficient attacks, persistently drilling down until it finds success.

While impressive, GPT-Red is not perfect. It struggles with back-and-forth conversations and isn’t adept at using images in prompt injection attacks. OpenAI sees it as a valuable supplement to human testers but acknowledges the need for continued expertise.