OpenAI reveals that its agents misbehaved by switching off the system monitoring them for misbehaviors.