Security Filter

Find it and protect it before it goes to an AI

It spots personal data and internal keywords before they go to an AI. On a hit your organization decides: send with a mask, send with a warning, or stop the send. The setting applies to everyone and individuals cannot turn it off.

From sending to the reply

Detect

It scans what is about to be sent for built-in kinds and your organization's own rules.

Apply the chosen action

Mask, warn, or stop the send, as chosen for each kind and each rule.

Send to the AI

With the mask action, the text goes to the AI with placeholders. With warn and send, the value goes as it is. With stop, nothing is sent.

Restore in the reply

If the reply contains a placeholder, it is turned back into the real value only when shown on screen.

Three actions

Chosen for each built-in kind and each custom rule.

Mask and send

The value is replaced with a placeholder such as "Email_1" before sending. If the reply contains the placeholder, the screen shows the original value again.

Warn and send

The send is not stopped, but the person is told that an organization rule was hit. This action does not mask the value.

Stop the send

Nothing goes to the AI. The screen says the message was not sent because of an organization rule, and no PFC is used.

What it covers

On Enterprise you can switch the items other than the current message on or off. By default all are on. Some paths check only the current message, and some check everything regardless of the switches. The contents of images, audio and video are not checked on any path.

  • check
    The current message

    Always on and cannot be switched off.

  • check
    Conversation history

    Content from earlier turns is checked too.

  • check
    Attachments

    Text that can be read from attached documents is checked. Text inside an image you attach is not.

  • check
    Knowledge (RAG)

    Content pulled from knowledge before it is passed to an AI is checked.

  • check
    Web page text from a URL

    Text fetched from a URL is checked.

  • check
    Prompts for image and other generation

    The instruction text for generating images and similar is checked.

What it finds out of the box

Email addresses, phone numbers, postal codes, Japanese My Number, credit card numbers, bank account numbers, addresses (in the form of a Japanese prefecture through a street number), API keys and US Social Security numbers (SSN). On Enterprise you choose, for each kind, whether it is detected and which action applies.

Your own rules

Add words that only your company uses, such as project names or client names. Keywords (several words, with optional fuzzy matching) and regular expressions are supported. Each rule has its own action, and you can name the placeholder (for example "Client"). Each organization can have up to 20 regular-expression rules.

Example

If someone sends "Write a reply about Mr. Tanaka (tanaka@example.com / 090-1234-5678)", the email address and phone number are replaced and the AI receives "...Mr. Tanaka (<Email_1> / <Phone_1>)...". If the reply contains a placeholder, the screen shows the original value again.

Personal names such as "Mr. Tanaka" are not found out of the box. Cover names and internal terms with your own rules.

What each plan includes

項目BusinessBusiness+Enterprise
Mask before sending, restore in the replyYesYesYes
Per-kind on/off and action——Yes
Switching what is covered (history, attachments and so on)——Yes
Your own rules——Yes
Standard auto-mask that individuals cannot turn offYesYesYes

What goes in the audit log

Only the kind and number of hits, the action (mask, warn, stop) and the rule ID are recorded. The matched values themselves, the rule contents and the conversation text are not.

Limits to know about

  • It finds things that fit fixed patterns and the rules your organization adds. It does not guarantee that every piece of sensitive information is found.
  • Personal names are not detected out of the box. Phone number detection centers on numbers of 10 to 11 digits.
  • What else is covered, such as history and attachments, depends on the path. Some paths check only the current message.
  • The contents of images, audio and video are not checked. Text shown in an image (for example an email address in a screenshot or a photo of a business card) is not found and goes to the AI as it is.
  • Content sent for voice-input transcription, knowledge ingestion (reading text from images, speech recognition and embeddings) and content-safety checks is not masked.
  • Only items set to "Mask and send" keep the original value from the AI. "Warn and send" does not mask the value, so the value goes to the AI.
  • Only an Enterprise organization's admins can change the per-kind actions and custom rules.

How to set it up

01

Open Security Filter

It is under Settings and is shown only to organization admins.

02

Choose an action for each kind

For each kind, such as email addresses, switch it on or off and choose mask, warn or stop.

03

Add your own rules

Add your own terms as keywords or regular expressions and choose an action.

04

Check it

From a member account, send a phrase that should hit. A stop rule shows that the message was not sent because of an organization rule.

Enterprise is available online: choose your seats and shared PFC, then pay by card.

Go to the sign-up page
Language日本語English