Labs at least must study prompts in an airgapped fashion. From there, consider how they could generate synthetic data to train another model. After, require trusted staff to do multiple levels of independent granular reviews of all fruits of the highest-value stolen inputs. (Or for model training data only, data never has to leave the airgap.)
Definitely risky, anyway. Surely some AI user has sent data, in confidential mode, with a unique shape they expect to be able to recognize if a later model recreated a facsimile even with heavy substitutions… but labs could bring risk of getting caught (over next few years) down quite low with extraordinarily ultraparanoid strategy. (But hopefully everybody is just behaving!)