Moon Blur protects what goes out to the AI model, and only the data it detects. Some content doesn’t go through it, its detection can miss a value, and a message that is too long is only analyzed in part.
What Moon Blur doesn’t maskLink to the What Moon Blur doesn’t mask section
| Content | Why |
|---|---|
| Data sent to your connected services (connectors, MCP servers, email inbox) | When the AI acts for you (sending an email, creating a ticket, adding an event), the real values are put back into what the service receives, otherwise the action would make no sense. |
| Images and photos | They are passed as is to the models that read images: an address or a number visible in a photo is not masked. |
| Voice | The audio of dictation and read aloud is not masked. |
| Web searches | The query sent to the web search engine goes out unmasked. |
| Descriptions of images and videos to create | The text describing the image or video to generate goes out as is. |
| The names of your email inboxes and MCP servers | The names you give them go out as is. |
| A page published through a public link | Before publishing, Moon Blur flags the personal data the page seems to contain, but the published page keeps the real values. The images on the page are not analyzed. |
In a confidential chat, a scanned document is attached as images, not as a PDF: its content is then not masked (Scan a document).
Detection that can miss thingsLink to the Detection that can miss things section
Detection is automatic. It can miss a value:
- Names of people, clients, companies or projects are not spotted on their own. Add the ones that matter to Your own data to mask.
- Unusual formatting: a number split differently from usual, without its label, or an unexpected abbreviation may not be recognized. Many numbers (case, contract, customer) are only spotted next to their label.
- Code and web addresses: automatic detection doesn’t scan text placed in a code block or between backticks, nor web addresses. An email address in a code snippet goes through as is. The terms on your personal list, however, are replaced there too.
- Context: even when a value is masked, sensitive information can be inferred from the rest of the message.
The details below each message (“Show”) show what was masked: check them when a message contains important data. The list of recognized types is in Detected data.
Partial analysis of a very long textLink to the Partial analysis of a very long text section
Moon Blur analyzes up to 500,000 characters of a single text. Beyond that:
- only the beginning is analyzed by automatic detection; the rest goes out without that masking;
- the terms on your personal list are still replaced in the whole text;
- the badge below the message turns amber, with “partial analysis” and the note “Very long message. Only the beginning was analyzed.”
When it is an older message in the conversation that was too long, the badge says “An earlier message in this conversation was only partially analyzed.” and offers “New chat” to start again from a short history.
To have a very long text analyzed in full, split it into several messages.
If the analysis failsLink to the If the analysis fails section
When Moon Blur can’t analyze a message, the message is not sent to the model. A message tells you so:
| What the message says | What to do |
|---|---|
| Moon Blur couldn’t analyze the message and nothing was sent. | Send the message again; if the error comes back, open a new chat. |
| Moon Blur can’t read your list of data to mask right now. | Wait a few moments and send the message again. |
| Moon Blur can no longer read some terms on your list of data to mask, so they couldn’t be masked. | Delete the terms marked “Unreadable term” in your list, now in “Settings” › “Data and privacy” › “Moon Blur”, then add them again. |
As a last resort, the platform has an emergency setting that lets an unanalyzed message go through. That message then carries the red badge “Moon Blur · sent without protection”. This setting never applies when your personal list is in use: a message is never sent with your terms in clear text.
In the API and Moon CLILink to the In the API and Moon CLI section
The developer API and Moon CLI follow their own rules, described in Privacy and Moon Blur in the API and Moon Blur in Moon CLI.
Before you sendLink to the Before you send section
- Remove information the request doesn’t need.
- Check what you are allowed to share.
- Avoid sending secrets and login credentials, even masked.
- Add the client and project names you often mention to your list.
- Read the answer over before using it in your work.
For Moon AI’s commitments on how data is handled, see Security and commitments.