The Microsoft AI research division accidentally leaked dozens of terabytes of sensitive data starting in July 2020 while contributing open-source AI learning models to a public GitHub repository.