GPT-Red scales prompt-injection attacks through self-play
OpenAI trained an internal attacker alongside defender models across controlled tool environments. It used generated attacks in GPT-5.6 training and reports six times fewer failures on its hardest direct prompt-injection benchmark than its best model four months earlier.
Why it matters: Automated adversaries can generate attack diversity and volume that human exercises cannot, making security testing part of model training rather than a one-time release gate.