“In addition to the ML models themselves, the exposed data can include training datasets, hyperparameters, and sometimes even raw data used to build models,” a security researcher said.