Those accounts then undertake more coordinated attempts to "jailbreak" or otherwise expose information about AI models that is not supposed to be made public, which is saved and applied to their own AI model building and training.