OpenAI uses internal AI red team to defend GPT-5.6

OpenAI has developed and deployed an internal artificial intelligence system, named GPTRed, that autonomously attacks its own models to discover and patch security vulnerabilities before they can be exploited. The company detailed the system this week, revealing it was instrumental in training its latest GPT-5.6 model family, significantly improving its resistance to prompt injection attacks.

GPTRed operates as an automated red teamer, using self-play reinforcement learning to generate increasingly sophisticated prompt injection attacks against a set of defender models. In this adversarial training loop, GPTRed earns rewards for successful exploits, while the defenders are rewarded for resisting them and completing their original tasks. This process forces both sides to continuously improve, with GPTRed discovering novel attack methods as the defenders harden. OpenAI stated it dedicated computational resources on the scale of its largest post-training jobs to this safety effort, an unprecedented investment for such work.

The system has demonstrated a high degree of effectiveness. In one internal evaluation replicating a 2025 indirect prompt injection test, GPTRed succeeded in 84...