Skip to content

Personal-data masking: what you notice

What it does

When you chat with an external model on Mid Privacy, masking replaces names, numbers and addresses in your message with placeholders before the AI processes them. The model does not see the values that were masked. Where the answer repeats a placeholder, you read the real value again. On Privacy Off nothing is masked.

What you see of this in the composer and in replies:

  • The Privacy pill in the composer (on a phone: the privacy tile before the microphone, which opens a Privacy sheet with the pill) shows the state for the next message. Full Privacy means the model is a local Myra model, so your message stays at Myra and is not masked. With masking on, only data sent to web search and tool calls is masked. Mid Privacy means masking is on for an external model. Privacy Off means your data goes to the external model unmasked, and the pill shows an open lock.
  • The Preview masking button shows a draft as it will be masked, before you send it.
  • A chip on a reply, PII masked: followed by the types, tells you what was replaced in that message.

How to ask

You write your message as usual. To check what masking does to a draft, enter the draft in the composer and click Preview masking. The dialog opens and runs the detectors.

  1. Read the masked items. Each one shows its type and a number, counted per type, for example Person 1 and Email 1. A second, different person is Person 2.
  2. If you want an item to go out as it is, click it. This sends that item unmasked.
  3. If the detectors missed something sensitive, select the text and click Mask selected text.
  4. Click Looks good — send to send the draft with the masking you see.

To change the route for the next message, click the Privacy pill. It switches between Mid Privacy and Privacy Off, if a protected alternative is available. If the pill is not a button, no alternative is available.

You can ask the assistant "Is masking active?". It answers from the policy for the turn. That is not a statement that your message contained personal data.

Tips

  • In your first seven days, a one-line hint under the input explains the pill: "If "Mid Privacy" is selected, we replace names, numbers, and addresses with placeholders before the AI processes your input."
  • On a gateway with PII protection, the preview is required before the first send.
  • If something in a message was masked and the message had attachments, the PII masked chip also shows Files in this message (n). Click it to see what happened to each file.

Limits

  • Preview masking does not appear for a local Myra model, because nothing is masked on the way in.
  • In Privacy Off, names, emails and IDs leave Myra unmasked. A banner says so before you send.
  • If your organization enforces masking, the pill is read-only with a closed lock and masking cannot be turned off. The preview offers no unmask action.
  • Masking works on detected data. Text the detectors miss goes out as written unless you mask it by hand.
  • A placeholder that the model mangles or cuts off cannot be restored in the answer. It is shown as a typed placeholder instead. If the model paraphrases a value instead of repeating the placeholder, the real value is not put back at that spot.

When something goes wrong

  • The preview says "⛔ This message will be rejected". A guardrail blocks the draft and the dialog names it. Edit the draft, or contact your administrator if you think it is a false positive.
  • The preview says "PII detection incomplete". A detection service could not be reached, so the preview may not show all personal data. Review the text before you send.
  • The preview says the usage budget is used up, the trial has ended or a spending limit is reached. The dialog names the refusal after "The masking preview could not be loaded" and offers no Retry, because trying again cannot help. The same refusal would stop you from sending the message. An administrator can raise the budget or limit, and your billing contact can top up the balance. A short-lived refusal such as a busy workspace keeps Retry.
  • The preview asks you to type a message. The draft is empty. Close the dialog, enter the text and open the preview again.
  • A banner says "Data goes to a third-party provider without PII protection." You are on Privacy Off. If a protected route can run the model, the banner offers Protect automatically. Click it to switch to that route.
  • A notice says "Stopped by a protection rule of your organization". A rule of your organization found content in your message, in an attached file or in the answer to it that it does not allow to be processed here. When one of the built-in data-protection rules found it, the notice says what kind of data it is: a credit card number, a bank account number, an ID number, a tax ID or a health insurance number. The notice names the attached files. Remove that content from the file or the message, or ask differently, and send again, or ask your administrator if you think it is a mistake. Another model does not help for this notice.
  • A notice titled "Blocked before it left your workspace" says a file "cannot be checked for personal data". A PDF, image or audio file cannot be checked for personal data, so nothing was sent. Choose a model hosted by Myra or start a new chat with one, or paste the text into the message instead of attaching the file.
  • The pill does not switch. No protected alternative can run the selected model. Pick another model to change the route.