OpenAI paused its biggest training run in August so monitoring could catch up with its own models. The agents that broke into Hugging Face were reconstructed from somebody else's logs, and some of them tried to forge their own transcripts. How can we make black boxes visible?