OpenAI's Self-Play Red Team Flywheel

imjustnewatai · x · 2026-07-16

This piece discusses how OpenAI seems to have demonstrated a language model self-play loop similar to AlphaZero:

GPT-Red is trained via self-play: one side attacks while another group of models defends.
In new prompt injection scenarios targeting GPT-5.1, GPT-Red achieved an 84% successful attack rate, compared to only 13% by human red teamers.
OpenAI then used the attack samples found by GPT-Red to train GPT-5.6 Sol, which the author claims reduced the failure rate against GPT-Red's direct prompt injections to 0.05%.

The author speculates that this "attack-patch-attack" flywheel could eventually extend from post-training to pre-training: models participating in generating and validating training data, designing harder curricula, optimizing the training process, and even helping build next-generation models.

However, he also cautions:

It is not yet "models training themselves"; humans still set the goals, environments, and scoring criteria.
OpenAI also believes GPT-5.6 has not yet reached its defined "High" self-improvement threshold.
Uncontrolled synthetic data loops could reduce diversity and actually harm the model.

Original post →

OpenAI's Self-Play Red Team Flywheel

More from Models