They're supposed to be focused on genuine training improvements over existing data, including via RL. Gathering data under the table is misaligned behavior.