Jul 16, 2026
•
7 min read
GPT-Red Breaks Every Model It Faces — Then Makes the Next One Unbreakable
OpenAI has trained GPT-Red, an automated red-teaming model that successfully attacks every production model up to GPT-5.5 — then uses those attacks to make GPT-5.6 Sol 20x more resistant to prompt injections.
Ab
Aby Varghese