Models in separate experiments used the channel to exchange exploits as they repeatedly compromised OpenAI systems.