The latest Data Phoenix news tagged model-extraction.
OpenAI says editing an earlier message can make the model redo work because its API drops the private reasoning that followed that message.