OpenAI Uses AI Red Team to Strengthen GPT-5.6 Against Prompt Injection Attacks

According to OpenAI, GPT-Red operates through adversarial self-play reinforcement learning. In this process, AI agents generate increasingly sophisticated prompt injection attacks, while defender models learn to resist these threats. The successful attacks identified by GPT-Red are then incorporated into the training of subsequent defender models, creating a continuous improvement cycle.

OpenAI reported that GPT-Red achieved an 84% success rate in internal evaluation scenarios for GPT-5.6, a notable improvement compared to the 13% success rate achieved by human red teamers in the same tests. The company stated that GPT-Red is intended to scale AI safety efforts, complementing human red teamers and other existing security measures.

This development reflects a broader trend of utilizing AI to enhance AI security. Earlier this month, the Ethereum Foundation announced the deployment of AI agents to red-team its network infrastructure. OpenAI views GPT-Red as an internal tool due to its offensive capabilities, aiming to foster a cycle where current models contribute to making future models more secure, aligned, and trustworthy.