Masking: paste, attach, send and keep context
How to set, under Policy > Masking, what the extension does when people paste, attach or send, who gets Keep context and which document formats your team can use.
Under Policy > Masking you decide what the extension does when an employee pastes text, attaches a file or sends a message, and whether they see the real values again in the AI tool's answer. You choose a setting for each of these features separately. Manual masking, where the employee deals with a highlight themselves, is always available whatever setting you choose. What the employee sees in the extension is described in Masking when you paste, attach or send.
Masking in the dashboardThe extension setting for each feature, and whether a tool has its own setting.The four settings for each feature
For each feature you choose one of four settings under In the extension. If your organisation has never changed the setting, every feature is set to Available for everyone.
| Setting | What the employee sees |
|---|---|
| Off | Nothing, because the feature is not in the extension. |
| Available | A switch that is off, which they can turn on themselves. |
| On by default | A switch that is on, which they can turn off. If they made a choice of their own earlier, that choice is kept. |
| Always on | A switch locked in the on position, with the text "Set by your organisation". |
The employee finds the switches in the extension under When to mask and Keep context.
Mask on paste
Transcript
On screen: Heading: "Sensitive data in pasted text". A browser window with ChatGPT at chatgpt.com; the BeeSensible extension icon sits on the right of the address bar. In the conversation the user asks: "I have notes from a customer call. Can you summarize them?" ChatGPT replies: "Sure. Paste the notes below." Next to it a document "Customer call notes" appears with the text: "Lucía Ortega called about her refund to DE89370400440532013000. Her passport number is NV4837291. She can be reached at +31 6 12345678." The text is selected and the label "Copied" appears.
Sensitive data often ends up in an AI tool because you paste text into it, such as part of an email or a customer file.
On screen: Heading: "When an AI tool receives the text", with the label "The risk". The text is pasted into the ChatGPT text field. The name and the phone number get a yellow line underneath, the IBAN and the passport number a red one. At the same time the four details fly from the text field to the address bar, with the label "Already on its way to the AI tool". Next to the send button it says "Send not clicked yet".
Some AI tools already send what is in the text field before you click send, so masking it afterwards comes too late.
On screen: Heading: "Checking before pasting", with the label "When to mask". The same text is pasted again. The text field stays empty and a BeeSensible panel appears above it: "BeeSensible is checking what you paste", with "Esc to cancel". The panel then shows "4 sensitive details found in what you paste" and "Nothing is in the text field yet. This is how it goes in when you paste masked." Below that is the text with placeholders: "[NAME1] called about her refund to [IBAN1]. Her passport number is [PASSPORT1]. She can be reached at [PHONE1]."
With masking when pasting, BeeSensible first checks what you paste, and only then does the text go into the text field.
On screen: Heading: "Pasting masked or the original", with the label "When pasting". The view zooms in on the panel. At the bottom are the buttons "Paste masked", "Paste original" and "Cancel", and a checkbox "From now on paste masked right away". The pointer moves over "Paste masked" and "Paste original".
Above the text field you see what was found, with a placeholder in place of each detail, and you choose whether to paste the masked text or the original.
On screen: Heading: "Pasting masked straight away", with the label "Paste right away". The checkbox "From now on paste masked right away" is ticked, and then "Paste masked" is clicked. The panel closes and the text field now holds the text with the placeholders.
If you do not need that step, tick the box to paste masked text straight away from now on.
On screen: Heading: "What the AI tool receives", with the label "What it changes". The send button is clicked. The four placeholders fly to the address bar, with the label "This is what the AI tool gets". The new message in the conversation shows [NAME1], [IBAN1], [PASSPORT1] and [PHONE1] instead of the details.
When you send the message, only the placeholders go to the AI tool.
On screen: End card: the BeeSensible logo with the title "Masking when you paste".
With mask on paste, pasted text goes through detection before it reaches the AI tool's input field. If detection finds sensitive data, the employee sees what was found and chooses Paste masked, Paste original or Cancel. If detection finds nothing, the text is pasted as it was. If detection does not respond, the employee decides, because BeeSensible never pastes unchecked text of its own accord. Formatting is removed from pasted text in every case. The feature works in ChatGPT, Claude, Gemini and Copilot on copilot.com.
Under When pasting you choose how much room the employee has:
| Choice | What happens |
|---|---|
| User chooses | The employee decides each time, and can skip the check from then on with "From now on paste masked right away". |
| Always show the check | The check appears every time, so the employee always sees what was found. |
| Always paste masked | Whatever is found goes into the field masked, without a check. Pasting the original is then not possible. |
Paste has a feature of its own because ChatGPT sends the draft while someone is still typing. Pasted text can therefore be on its way before the employee presses send, which means masking on send comes too late for that text.
Mask on attach and on send
With mask on attach, the extension holds back a file that someone adds in ChatGPT, Claude or Copilot (copilot.com) before the AI tool sees it. The employee then chooses Check and mask, Add the original or Cancel. If they choose to check it, the file opens in the sidebar, and after the check they add the anonymised copy to the conversation. This feature needs the Document Redaction module. File types that BeeSensible cannot check pass through unchecked, and attaching does not work in Gemini.
Transcript
On screen: Heading: "Highlights as you type". A browser window with ChatGPT. In the conversation the user asks: "Can you help me with a customer email?" ChatGPT replies: "Of course. What should it say?" Someone types in the text field: "Rewrite this as a customer email: Lucía Ortega, passport NV4837291, gets her allowance from next month. Questions? Call +31 6 12345678." The name and the phone number get a yellow line underneath, the passport number a red double line. The BeeSensible tab on the right edge of the window shows the number of highlights, 3.
BeeSensible marks sensitive data as you type. Masking every detail yourself takes time, and it is easy to press Enter before you have checked everything.
On screen: Heading: "Turning on When sending", with the label "When to mask". The text field is cleared. The BeeSensible tab is clicked and the side panel opens on the "Privacy" tab, showing "No sensitive items detected". Then the gear icon is clicked and "Settings" opens. The panel scrolls to the section "When to mask": "You can always mask by hand. Below you choose when BeeSensible does it for you." Below that are the switches "When pasting" and "When sending". Under "When sending" it says: "Everything still detected becomes a placeholder the moment you send. After that your message goes through." That switch is turned on and the panel is closed.
You can leave the masking to BeeSensible instead. In the settings of the side panel, under When to mask, turn on When sending.
On screen: Heading: "What happens when you send", with the label "When sending". The same message is typed again and Enter is pressed. Above the text field "BeeSensible is checking your message" appears briefly. Then the field shows [NAME1], [PASSPORT1] and [PHONE1] in place of the details, with a dotted line underneath, and the message is sent.
You then type as usual. When you press Enter, BeeSensible checks your message and turns everything it finds into a placeholder before the message is sent.
On screen: Heading: "What was masked", with the label "What it changes". The view zooms in on the sent message with the placeholders, with the label "3 details masked" above it. ChatGPT replies: "Dear [NAME1], Good news: your allowance starts next month. Any questions? Call us on [PHONE1]."
The sent message shows which details were masked.
On screen: End card: the BeeSensible logo with the title "Masking when you send".
With mask on send, the extension turns everything that is still highlighted into a placeholder when the employee presses Enter or the send button, and then sends the message. If the extension is still checking the last keystrokes, it waits up to two and a half seconds and then sends the message as it stands. It does not automatically mask a language such as "English", because in a prompt that is usually part of the instruction. In the figures, these masks count as masked.
Keep context
Transcript
On screen: Heading: "Placeholders in the answer". A browser window with a ChatGPT conversation. The user asks: "Summarise the note about [NAME1] for the mentor meeting." ChatGPT replies: "[NAME1] was late four times in the past three weeks and missed two tests. Mother [NAME2] can be reached on [PHONE1]."
Sensitive details often give an AI tool the context it needs, such as the names in a report about a student. If you mask them, the answer is full of placeholders.
On screen: Heading: "Turning on Keep context", with the label "Settings". The window darkens and an enlarged setting from the extension appears: the switch "Keep context" with the label "Beta". Below it: "Masked details are back in place in the answer, on your screen only. When you mask something in a prompt, BeeSensible remembers which placeholder stands for which value, on this device only. The AI tool only ever sees the placeholders." The switch is turned on. A line "Works in" appears with ticks next to ChatGPT, Claude, Gemini and Copilot, followed by the note: "Restored details are marked with a dotted line. Conversations you share or export from the AI tool keep the placeholders."
With Keep context, BeeSensible puts the real details back in the answer. You turn it on once in the settings.
On screen: Heading: "The real details in the answer", with the label "In the answer". Someone types in the text field: "Email Femke Brouwer: the mentor meeting about Esmee van der Wal moves to Thursday. Call her on +31 6 12345678." The two names and the phone number get a line underneath. After Enter they become [NAME1-K7], [NAME2-K7] and [PHONE1-K7] and the message is sent. The sent message and the ChatGPT answer show the real names and number, underlined with a dotted line: "Dear Femke Brouwer, The mentor meeting about Esmee van der Wal moves to Thursday. If that does not suit you, we will call you on +31 6 12345678." Next to it a card "What ChatGPT gets" shows the same answer with the placeholders, with the label "On this device only".
You mask as you normally would, and the AI tool only receives the placeholders. In the answer you read the real names, underlined with a dotted line. BeeSensible keeps track of which placeholder belongs to which detail on your device only.
On screen: Heading: "Switching and copying", with the label "Conversations". The ChatGPT sidebar opens with the chats "Mentor meeting email", "Lisbon day plan", "Quarterly figures summary" and "Team outing ideas". The view switches to "Lisbon day plan" and then back to "Mentor meeting email"; the real details are still there. The answer is copied, and a card "Pasted into your email" shows the text with the real names and phone number.
The real details stay in place when you switch to another conversation and come back, and when you copy the answer.
On screen: Heading: "The AI tools it works in", with the label "Works in". The window darkens and the logos of ChatGPT, Claude, Gemini and Copilot appear side by side.
Keep context works in ChatGPT, Claude, Gemini and Copilot on copilot.com, and you choose for each tool whether it is on.
On screen: End card: the BeeSensible logo with the title "Keep context".
With Keep context, the extension puts the masked values back into the AI tool's answer, only on the employee's screen and with a dotted underline. The AI tool itself only ever sees the placeholders. Besides the setting, you make two choices for this feature:
Keep context
| Setting | Choice |
|---|---|
| Who gets it | Admins only, so you can try the feature yourself first, or Everyone. |
| Memory | Browser session only: the memory is cleared when the browser closes, on sign-out or after a day. 30 days: the memory is stored encrypted on the device and is cleared after 30 days without use or on sign-out. |
The memory, which records which placeholder stands for which value, stays on the device and never reaches BeeSensible. Keep context works in ChatGPT, Claude, Gemini and Copilot on copilot.com, but not in Microsoft 365 Copilot and not in the desktop app.
File formats and per-tool settings
With the More file formats for documents setting, your team can anonymise Office, OpenDocument and text files, scans and images as well as PDFs. Under For your organisation, choose On, which is the default, or Off, after which only PDFs are possible. The formats are listed in Documents.
For ChatGPT, Claude and Gemini you can deviate from the organisation's policy. You set this on the tool's Policy tab in the AI register, as described in Policy per tool. When a tool deviates, the feature on this page shows "Deviates for:" followed by the names of those tools. Who gets Keep context is set for the whole organisation only.
The audit log records every change to these settings, and the extension picks up new settings within five minutes. Which data is masked follows from the app's detection profile.
