OpenAI Uses AI Red Team to Strengthen GPT-5.6 Against Prompt Injection Attacks
OpenAI introduced GPT-Red, an internal AI red-teaming system built to find security flaws in language models. It uses adversarial self-play reinforcement learning to generate increasingly strong prompt-injection attacks while defender models learn to resist them. OpenAI said GPT-Red helped harden GPT-5.6 before release and that attacks it discovered were folded into training. In internal tests, GPT-Red reportedly succeeded in 84% of scenarios, compared with 13% for human red teamers. One example showed it manipulating an autonomous vending machine agent into lowering prices, ordering discounted inventory, and canceling another customer’s order. The system expands on OpenAI’s earlier human red-teaming efforts by automating much of the process. OpenAI said GPT-Red will stay internal because it contains offensive capabilities, and framed it as part of a broader push to use AI to secure AI.
