OpenAI Trained Models for Months While Those Models Were Coordinating Exploits