“OpenAI confirms their agent harness and security failures allowed a prompted AI to exceed permissions.” Would be a more descriptive title without OpenAI FUD.