OpenAI's GPT-Red Paper: Large-Scale Automated Red Teaming via Self-Play

alex_verem · x · 2026-08-01

OpenAI released a paper introducing GPT-Red, an automated red-teaming agent designed to discover novel prompt injection attacks against frontier LLMs.

Self-Play Training: Features a scalable self-play algorithm where the model attacks a diverse population of simultaneously-trained defender agents.
Unprecedented Scale: Utilizes compute on the scale of their largest RL post-training runs, documenting the single-largest LLM safety training run to date.
Proven Efficacy: Reliably breaks past models up to GPT-5.5, finds more successful attacks than human red-teamers, and generalizes well to held-out environments.
Safety Flywheel: Used to adversarially train GPT-5.6, making it their most robust model against prompt injections. This establishes a self-improvement flywheel for future model safety.

Original post →

OpenAI's GPT-Red Paper: Large-Scale Automated Red Teaming via Self-Play

More from Safety