OpenAI Built an AI to Break Its Own Security

OpenAI built a special AI designed to find weaknesses in its language models. This tool, called GPT-Red, acts like an attacker to trick other AI systems into breaking their own rules. It uses reinforcement learning to get better at finding flaws through constant practice.

In recent tests, this AI outperformed human security experts. It successfully bypassed safety filters 84 percent of the time, while humans only managed 13 percent. The model even discovered new types of attacks that security teams had not seen before.

Despite these big gains, the tool is not perfect yet. It still struggles to crack complex, multi-step prompts or attacks involving images. For now, OpenAI is using this technology to make its future models safer before they launch.