I am not sure what could be safer than failing on invalid data - if invalid data is encountered, failing would be better than e.g. guessing and silently corrupting data.