Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring