OpenAI details GPT-Red, an internal model built to attack other AI systems

What happened

OpenAI has described GPT-Red, an internal red-teaming model trained through self-play to generate adversarial attacks against other AI systems.

The facts in this brief are limited to what the named organizations published. Product capabilities, research results, and future plans are attributed to those sources and should be read as their claims unless independently reproduced.

The useful context

OpenAI says the research model automates adversarial testing and helped improve GPT-5.6 defenses against prompt injection. GPT-Red remains separate from deployed models.

What to watch

Watch for primary documentation, independent evaluations, implementation details, and any update that changes the availability or scope described here. AI Wire will date material changes and keep corrections attached to this record.

Sources & evidence