Indeed - if you need to be absolutely confident about security and privacy, run a model locally and audit the inference software and potential tooling.