OpenAI has introduced GPT-Red, an automated red-teaming system that finds vulnerabilities in its language models before release. The company said the tool was used to train GPT-5.6, reducing failures on one of its hardest prompt injection benchmarks. GPT-Red succeeded in 84% of internal evaluation scenarios, compared with 13% for human red teamers in the same tests.
How GPT-Red Works
Named after cybersecurity red teaming, GPT-Red is trained through self-play reinforcement learning. It generates progressively stronger prompt injection attacks while defender models learn to resist them. OpenAI said every successful attack is used to improve defenders, pushing GPT-Red to find broader and more complex failures. In one case study, the system manipulated an autonomous vending machine agent into lowering prices and canceling orders before the vulnerabilities were disclosed and fixed.
Scaling Safety for Advanced Models
OpenAI wrote on X that as model capabilities grow, safety and alignment must scale with them, and red-teaming is essential but difficult to scale. GPT-Red is one way the company is addressing that bottleneck. The system follows OpenAI's 2023 Red Teaming Network, which recruited outside researchers to probe ChatGPT and other models. GPT-Red automates much of that process, using AI to generate adversarial tests at a scale human researchers alone could not achieve.
Broader Industry Shift
OpenAI's announcement reflects a broader trend of using AI to secure AI. Earlier this month, the Ethereum Foundation deployed AI agents to red-team critical network infrastructure, uncovering a vulnerability in Ethereum consensus client software. Researchers noted that AI agents can search larger codebases than humans, but the challenge has shifted from finding potential bugs to proving which ones are exploitable.
OpenAI said GPT-Red will remain an internal tool because it contains intentionally developed offensive capabilities. The company believes it has started to unlock a flywheel for safety, where today's models can be used to make tomorrow's models more robust, aligned, and trustworthy.