OpenAI reveals how to re-align a misbehaving GPT-4o model