Red teaming

Technique · Safety & Alignment · 1 item since Jul 2026

Attacking models on purpose to find unsafe behavior before release.

Organizations: OpenAI
Techniques: Self-play

Trends

Frontier labs slow down and publicly report their own risks

Items

Jul 15
OpenAI's GPT-Red automated red-teamer
Trained by self-play RL, the red-teaming model finds prompt-injection attacks on 84% of held-out scenarios (versus 13% for human red-teamers) and was used to adversarially train GPT-5.6 Sol to roughly sixfold fewer failures on OpenAI's hardest direct-injection benchmark.