Skip to main content
We would rather you know where the edges are.

​
Not yet supported

Cloak detects identifiers. It does not detect conditions, procedures, or drugs, and nothing is reported under the phi category. A medical record number is detected as an account id. Health data needs a model trained on clinical text and is on the roadmap.
Cloak works on text. For a voice agent, that means the transcript: protect the text of each turn before it reaches the model. Audio itself, and text inside images, are not read. Non-text parts pass through untouched and are reported as such. Audio and image protection are on the roadmap.
A streamed response is not restored automatically, because its chunks reach you before the call ends. Restore the text you assemble with guard.cloak.restore.

​
Limits today

A payload over the size limit is reported as not checked, and Cloak fails closed on it. Split very large payloads before submitting them.

​
Accuracy

Detection is a model, not a rule set, so it has false positives and false negatives.
  • Structured values such as emails, phone numbers, cards, and IBANs are the strongest.
  • Names can be over-reported. Tune minConfidence against your own traffic.
  • A value written in an unusual format may be missed.
Test Cloak against your own data before relying on it, and use report mode on the semantic pass to watch what it does.

​
Data handling

  • The text of a request is sent to the control plane for detection, over TLS. The response contains types and ranges only.
  • The control plane does not record request text for Cloak, and telemetry carries counts only.
  • Originals never leave your process. The vault is yours.
See Security and privacy.