notebook

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

JupyterLab's HTML sanitizer allowlists data-commandlinker-command and data-commandlinker-args on button elements, while CommandLinker listens for all click events on document.body and executes the named command without checking whether the element came from trusted JupyterLab UI. A notebook with a pre-saved HTML cell output containing a deceptive button can trigger arbitrary JupyterLab commands - including arbitrary code execution - on a single user click, without any code being submitted for execution by the user.

Impact

An attacker who shares a notebook or a Markdown file - via email, GitHub, or a Binder link - can invoke an arbitrary command upon a single click by the victim. The button can be rendered inside the output area and be visually indistinguishable from a legitimate widget. No kernel needs to start; the HTML output is stored in the notebook file and displayed immediately on open.

Single-click impact

An attacker convincing the victim to click on a single button or link can:

• execute arbitrary code in the available kernels,
• delete files leading to information loss; in principle the loss could be unrecoverable, depending on server configuration and attack complexity,
• open multiple kernels/terminals at once, or create multiple files at once, putting significant stress on the server and thus deny availability for other users when using standalone multi-tenant jupyter-server deployment, and to a lesser degree impact availability on JupyterHub deployments.

The arbitrary code execution will be immediately visible to the user; and can be halted by the timely user intervention. The deletion of files can be silent and go unnoticed for some time.

Multi-click attacks

An attacker who convinces the victim to click on multiple buttons in specific order and to grant access to clipboard (or in scenarios where the user already granted keyboard access) can obtain full access to the terminal and execute arbitrary commands in the environment with access scope that might exceed that of available kernels. Only users of Chromium-based browsers are susceptible to this expanded variant of the attack.

The execution of commands in the terminal would be immediately visible to the user.

Impact of third-party extensions

The impact described above assumes a plain JupyterLab/Notebook installation. In environments with frontend extensions that contribute additional commands the attack surface is increased by the functionality covered by these commands.

Patches

JupyterLab 4.5.7

Workarounds

No workarounds are available for end-users.

Downstream applications inheriting from JupyterFrontEnd or JupyterLab can effectively disable the CommandLinker by passing commandLinker: new CommandLinker({ commands: new CommandRegistry() }) option in the initialization options.

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-files

Impact

A stored Cross-Site Scripting (XSS) vulnerability in Jupyter Notebook allows attackers to steal authentication tokens from users who open malicious notebook files and interact with elements that the attacker can make look indistinguishable from legitimate controls (single click interaction).

The vulnerability enables complete account takeover through the Jupyter REST API, allowing the attacker to:

1. Read all files
2. Modify/create files
3. Access running kernels and execute arbitrary code
4. Create terminals for shell access

Patches

Jupyter Notebook 7.5.6 and JupyterLab 4.5.7 include patches for this vulnerability.

Workarounds

The help extension can be disabled via CLI:

jupyter labextension disable @jupyter-notebook/help-extension
jupyter labextension disable @jupyterlab/help-extension

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-output-and-files

Acknowledgments

Reported by Daniel Teixeira - NVIDIA AI Red Team

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

JupyterLab's HTML sanitizer allowlists data-commandlinker-command and data-commandlinker-args on button elements, while CommandLinker listens for all click events on document.body and executes the named command without checking whether the element came from trusted JupyterLab UI. A notebook with a pre-saved HTML cell output containing a deceptive button can trigger arbitrary JupyterLab commands - including arbitrary code execution - on a single user click, without any code being submitted for execution by the user.

Impact

An attacker who shares a notebook or a Markdown file - via email, GitHub, or a Binder link - can invoke an arbitrary command upon a single click by the victim. The button can be rendered inside the output area and be visually indistinguishable from a legitimate widget. No kernel needs to start; the HTML output is stored in the notebook file and displayed immediately on open.

Single-click impact

An attacker convincing the victim to click on a single button or link can:

• execute arbitrary code in the available kernels,
• delete files leading to information loss; in principle the loss could be unrecoverable, depending on server configuration and attack complexity,
• open multiple kernels/terminals at once, or create multiple files at once, putting significant stress on the server and thus deny availability for other users when using standalone multi-tenant jupyter-server deployment, and to a lesser degree impact availability on JupyterHub deployments.

The arbitrary code execution will be immediately visible to the user; and can be halted by the timely user intervention. The deletion of files can be silent and go unnoticed for some time.

Multi-click attacks

An attacker who convinces the victim to click on multiple buttons in specific order and to grant access to clipboard (or in scenarios where the user already granted keyboard access) can obtain full access to the terminal and execute arbitrary commands in the environment with access scope that might exceed that of available kernels. Only users of Chromium-based browsers are susceptible to this expanded variant of the attack.

The execution of commands in the terminal would be immediately visible to the user.

Impact of third-party extensions

The impact described above assumes a plain JupyterLab/Notebook installation. In environments with frontend extensions that contribute additional commands the attack surface is increased by the functionality covered by these commands.

Patches

JupyterLab 4.5.7

Workarounds

No workarounds are available for end-users.

Downstream applications inheriting from JupyterFrontEnd or JupyterLab can effectively disable the CommandLinker by passing commandLinker: new CommandLinker({ commands: new CommandRegistry() }) option in the initialization options.

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-files

Impact

A stored Cross-Site Scripting (XSS) vulnerability in Jupyter Notebook allows attackers to steal authentication tokens from users who open malicious notebook files and interact with elements that the attacker can make look indistinguishable from legitimate controls (single click interaction).

The vulnerability enables complete account takeover through the Jupyter REST API, allowing the attacker to:

1. Read all files
2. Modify/create files
3. Access running kernels and execute arbitrary code
4. Create terminals for shell access

Patches

Jupyter Notebook 7.5.6 and JupyterLab 4.5.7 include patches for this vulnerability.

Workarounds

The help extension can be disabled via CLI:

jupyter labextension disable @jupyter-notebook/help-extension
jupyter labextension disable @jupyterlab/help-extension

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-output-and-files

Acknowledgments

Reported by Daniel Teixeira - NVIDIA AI Red Team

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

JupyterLab's HTML sanitizer allowlists data-commandlinker-command and data-commandlinker-args on button elements, while CommandLinker listens for all click events on document.body and executes the named command without checking whether the element came from trusted JupyterLab UI. A notebook with a pre-saved HTML cell output containing a deceptive button can trigger arbitrary JupyterLab commands - including arbitrary code execution - on a single user click, without any code being submitted for execution by the user.

Impact

An attacker who shares a notebook or a Markdown file - via email, GitHub, or a Binder link - can invoke an arbitrary command upon a single click by the victim. The button can be rendered inside the output area and be visually indistinguishable from a legitimate widget. No kernel needs to start; the HTML output is stored in the notebook file and displayed immediately on open.

Single-click impact

An attacker convincing the victim to click on a single button or link can:

• execute arbitrary code in the available kernels,
• delete files leading to information loss; in principle the loss could be unrecoverable, depending on server configuration and attack complexity,
• open multiple kernels/terminals at once, or create multiple files at once, putting significant stress on the server and thus deny availability for other users when using standalone multi-tenant jupyter-server deployment, and to a lesser degree impact availability on JupyterHub deployments.

The arbitrary code execution will be immediately visible to the user; and can be halted by the timely user intervention. The deletion of files can be silent and go unnoticed for some time.

Multi-click attacks

An attacker who convinces the victim to click on multiple buttons in specific order and to grant access to clipboard (or in scenarios where the user already granted keyboard access) can obtain full access to the terminal and execute arbitrary commands in the environment with access scope that might exceed that of available kernels. Only users of Chromium-based browsers are susceptible to this expanded variant of the attack.

The execution of commands in the terminal would be immediately visible to the user.

Impact of third-party extensions

The impact described above assumes a plain JupyterLab/Notebook installation. In environments with frontend extensions that contribute additional commands the attack surface is increased by the functionality covered by these commands.

Patches

JupyterLab 4.5.7

Workarounds

No workarounds are available for end-users.

Downstream applications inheriting from JupyterFrontEnd or JupyterLab can effectively disable the CommandLinker by passing commandLinker: new CommandLinker({ commands: new CommandRegistry() }) option in the initialization options.

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-files

Impact

A stored Cross-Site Scripting (XSS) vulnerability in Jupyter Notebook allows attackers to steal authentication tokens from users who open malicious notebook files and interact with elements that the attacker can make look indistinguishable from legitimate controls (single click interaction).

The vulnerability enables complete account takeover through the Jupyter REST API, allowing the attacker to:

1. Read all files
2. Modify/create files
3. Access running kernels and execute arbitrary code
4. Create terminals for shell access

Patches

Jupyter Notebook 7.5.6 and JupyterLab 4.5.7 include patches for this vulnerability.

Workarounds

The help extension can be disabled via CLI:

jupyter labextension disable @jupyter-notebook/help-extension
jupyter labextension disable @jupyterlab/help-extension

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-output-and-files

Acknowledgments

Reported by Daniel Teixeira - NVIDIA AI Red Team

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

JupyterLab's HTML sanitizer allowlists data-commandlinker-command and data-commandlinker-args on button elements, while CommandLinker listens for all click events on document.body and executes the named command without checking whether the element came from trusted JupyterLab UI. A notebook with a pre-saved HTML cell output containing a deceptive button can trigger arbitrary JupyterLab commands - including arbitrary code execution - on a single user click, without any code being submitted for execution by the user.

Impact

An attacker who shares a notebook or a Markdown file - via email, GitHub, or a Binder link - can invoke an arbitrary command upon a single click by the victim. The button can be rendered inside the output area and be visually indistinguishable from a legitimate widget. No kernel needs to start; the HTML output is stored in the notebook file and displayed immediately on open.

Single-click impact

An attacker convincing the victim to click on a single button or link can:

• execute arbitrary code in the available kernels,
• delete files leading to information loss; in principle the loss could be unrecoverable, depending on server configuration and attack complexity,
• open multiple kernels/terminals at once, or create multiple files at once, putting significant stress on the server and thus deny availability for other users when using standalone multi-tenant jupyter-server deployment, and to a lesser degree impact availability on JupyterHub deployments.

The arbitrary code execution will be immediately visible to the user; and can be halted by the timely user intervention. The deletion of files can be silent and go unnoticed for some time.

Multi-click attacks

An attacker who convinces the victim to click on multiple buttons in specific order and to grant access to clipboard (or in scenarios where the user already granted keyboard access) can obtain full access to the terminal and execute arbitrary commands in the environment with access scope that might exceed that of available kernels. Only users of Chromium-based browsers are susceptible to this expanded variant of the attack.

The execution of commands in the terminal would be immediately visible to the user.

Impact of third-party extensions

The impact described above assumes a plain JupyterLab/Notebook installation. In environments with frontend extensions that contribute additional commands the attack surface is increased by the functionality covered by these commands.

Patches

JupyterLab 4.5.7

Workarounds

No workarounds are available for end-users.

Downstream applications inheriting from JupyterFrontEnd or JupyterLab can effectively disable the CommandLinker by passing commandLinker: new CommandLinker({ commands: new CommandRegistry() }) option in the initialization options.

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-files

Impact

A stored Cross-Site Scripting (XSS) vulnerability in Jupyter Notebook allows attackers to steal authentication tokens from users who open malicious notebook files and interact with elements that the attacker can make look indistinguishable from legitimate controls (single click interaction).

The vulnerability enables complete account takeover through the Jupyter REST API, allowing the attacker to:

1. Read all files
2. Modify/create files
3. Access running kernels and execute arbitrary code
4. Create terminals for shell access

Patches

Jupyter Notebook 7.5.6 and JupyterLab 4.5.7 include patches for this vulnerability.

Workarounds

The help extension can be disabled via CLI:

jupyter labextension disable @jupyter-notebook/help-extension
jupyter labextension disable @jupyterlab/help-extension

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-output-and-files

Acknowledgments

Reported by Daniel Teixeira - NVIDIA AI Red Team

Description

A Markdown file or a notebook output can carry a button that runs a JupyterLab command. Since 4.5.7 the first click on such a button opens a prompt naming the command, with Cancel, Run Once and Trust. Choosing Trust runs the command and marks the document trusted, so its later buttons run without a prompt.

Which document gets marked is decided when the user answers, not when they click. Both trust commands read whichever document is current at that moment: markdownviewer:trust uses tracker.currentWidget, and notebook:trust uses the equivalent lookup. If another document becomes current while the prompt is open, that document is the one marked trusted, and the one the user was looking at stays untrusted.

Users are affected if all of the following are true:

• they run JupyterLab 4.5.7 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• they open a Markdown file or notebook from someone else and click a button in it;
• they answer Trust rather than Cancel or Run Once; and
• a second document becomes the current one while the prompt is open, for example through a workspace restoring at startup, an extension activating a widget, or a command that opens a file

Impact

The user loses control over which document they trusted. Every button in the document that was (incorrectly) marked as trusted runs from then on without a prompt, so if that document also came from an untrusted source, its buttons fire on a single click with no confirmation. That is the state the prompt was added to prevent, in GHSA-mqcg-5x36-vfcg.

The command that ran had the user's consent, because the prompt named it and they pressed Trust. The document they were reading stays untrusted, so it prompts again on the next click.

Notebooks are written to disk

When the document marked is a notebook, the effect outlives the browser session. notebook:trust marks every cell trusted and then saves the file, which results in the server signing it, so the notebook stays trusted when it is next opened. The save also writes out any unsaved edits in that notebook. A second confirmation appears first, but it says only that a trusted notebook may execute hidden code, without naming the file, so it does not show the user which notebook is about to be trusted.

A Markdown preview marked this way stays trusted only until it is closed. Nothing is written to disk.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. The trust request now carries the document that was clicked, and the command is run only once that same document has been marked.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

Given the high complexity of a potential exploit, we advise treating this as a defense in-depth and update at earliest opportunity, rather than apply the workarounds listed below. However, for users running in highly sensitive environments who are unable to upgrade the workarounds are included below.

Turn off command linker, which removes the ability to execute frontend commands altogether:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

In such configuration the buttons in Markdown files and notebook outputs then do nothing when clicked. Documentation and extensions that use command links for navigation stop working, and the setting applies to every document, not only untrusted ones.

If disabling that setting is not desirable and update is not possible, advise users to choose "Run Once" rather than "Trust". "Run Once" makes the command run but no document is marked, so nothing can be trusted in error.

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

JupyterLab's HTML sanitizer allowlists data-commandlinker-command and data-commandlinker-args on button elements, while CommandLinker listens for all click events on document.body and executes the named command without checking whether the element came from trusted JupyterLab UI. A notebook with a pre-saved HTML cell output containing a deceptive button can trigger arbitrary JupyterLab commands - including arbitrary code execution - on a single user click, without any code being submitted for execution by the user.

Impact

An attacker who shares a notebook or a Markdown file - via email, GitHub, or a Binder link - can invoke an arbitrary command upon a single click by the victim. The button can be rendered inside the output area and be visually indistinguishable from a legitimate widget. No kernel needs to start; the HTML output is stored in the notebook file and displayed immediately on open.

Single-click impact

An attacker convincing the victim to click on a single button or link can:

• execute arbitrary code in the available kernels,
• delete files leading to information loss; in principle the loss could be unrecoverable, depending on server configuration and attack complexity,
• open multiple kernels/terminals at once, or create multiple files at once, putting significant stress on the server and thus deny availability for other users when using standalone multi-tenant jupyter-server deployment, and to a lesser degree impact availability on JupyterHub deployments.

The arbitrary code execution will be immediately visible to the user; and can be halted by the timely user intervention. The deletion of files can be silent and go unnoticed for some time.

Multi-click attacks

An attacker who convinces the victim to click on multiple buttons in specific order and to grant access to clipboard (or in scenarios where the user already granted keyboard access) can obtain full access to the terminal and execute arbitrary commands in the environment with access scope that might exceed that of available kernels. Only users of Chromium-based browsers are susceptible to this expanded variant of the attack.

The execution of commands in the terminal would be immediately visible to the user.

Impact of third-party extensions

The impact described above assumes a plain JupyterLab/Notebook installation. In environments with frontend extensions that contribute additional commands the attack surface is increased by the functionality covered by these commands.

Patches

JupyterLab 4.5.7

Workarounds

No workarounds are available for end-users.

Downstream applications inheriting from JupyterFrontEnd or JupyterLab can effectively disable the CommandLinker by passing commandLinker: new CommandLinker({ commands: new CommandRegistry() }) option in the initialization options.

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-files

Impact

A stored Cross-Site Scripting (XSS) vulnerability in Jupyter Notebook allows attackers to steal authentication tokens from users who open malicious notebook files and interact with elements that the attacker can make look indistinguishable from legitimate controls (single click interaction).

The vulnerability enables complete account takeover through the Jupyter REST API, allowing the attacker to:

1. Read all files
2. Modify/create files
3. Access running kernels and execute arbitrary code
4. Create terminals for shell access

Patches

Jupyter Notebook 7.5.6 and JupyterLab 4.5.7 include patches for this vulnerability.

Workarounds

The help extension can be disabled via CLI:

jupyter labextension disable @jupyter-notebook/help-extension
jupyter labextension disable @jupyterlab/help-extension

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-output-and-files

Acknowledgments

Reported by Daniel Teixeira - NVIDIA AI Red Team

Description

JupyterLab 4.5.0 enabled copying and pasing cells through the system clipboard. The paste path parses clipboard text as cell JSON and inserts the cells without clearing metadata.trusted, so a payload can declare its own output as trusted. JupyterLab does not sanitize a trusted output and evaluates the <script> elements it contains, so pasting the cell runs attacker's JavaScript in the JupyterLab origin. The same payload with "trusted": false is sanitized.

[
  {
    "cell_type": "code",
    "execution_count": null,
    "source": "",
    "metadata": { "trusted": true },
    "outputs": [
      {
        "output_type": "display_data",
        "metadata": {},
        "data": { "text/html": "<script>/* runs on paste */</script>" }
      }
    ]
  }
]

Users are affected if all of the following are true:

• they run JupyterLab 4.5.0 through 4.6.3, or an application that bundles it such as Notebook 7.5+;
• if running 4.5.2 or newer, useSystemClipboardForCells is enabled - it is off by default in 4.5.2+, and always enabled on 4.5.0 and 4.5.1 with no opt-out
• they run a Paste Cells command (e.g. from context menu, command palette, shortcut) while attacker-controlled text is in the clipboard; and
• pasteCodeCellsWithoutOutput is off, the default, so the pasted cell keeps its outputs

Impact

The script runs in the JupyterLab page, in the same origin and session as the authenticated user, so it can call the Jupyter Server REST API as that user: read and write any file under the server root, start a kernel and run code in it, and open a terminal where terminals are enabled. It runs on paste, without the user executing anything.

Only the output of a code cell can run a script this way. Markdown and raw cells are always sanitized, as the trust documentation states.

The attacker needs no JupyterLab account and no access to the server, only the user's clipboard at the moment of the paste. A web page can replace the text you copy from it, or write the clipboard directly once you have given it permission. Both need you to click or press a key on that page first.

Patches

JupyterLab v4.6.4 and v4.5.11 contain the patch. Cells read from the system clipboard that cannot be traced to the current window no longer carry metadata.trusted into the notebook, so their outputs are sanitized like any other content of unknown origin.

Users of applications that depend on JupyterLab, such as Notebook v7+, should update jupyterlab package too.

Workarounds

On 4.5.2 and later, confirm that the system clipboard is not in use for cells. The setting is off by default; check Settings, Notebook, or the raw value:

{
  "@jupyterlab/notebook-extension:tracker": {
    "useSystemClipboardForCells": false
  }
}

With that setting off, cell copy and paste use JupyterLab's in-application clipboard, which never leaves the page. The cost is that cells can no longer be copied between browser tabs or windows.

On 4.5.0 and 4.5.1 there is no such setting, and every paste reads the system clipboard and upgrade is highly recommended.

Pasting cells without their outputs also stops the script from running, at the cost of discarding outputs on every paste:

{
  "@jupyterlab/notebook-extension:tracker": {
    "pasteCodeCellsWithoutOutput": true
  }
}

JupyterLab's HTML sanitizer allowlists data-commandlinker-command and data-commandlinker-args on button elements, while CommandLinker listens for all click events on document.body and executes the named command without checking whether the element came from trusted JupyterLab UI. A notebook with a pre-saved HTML cell output containing a deceptive button can trigger arbitrary JupyterLab commands - including arbitrary code execution - on a single user click, without any code being submitted for execution by the user.

Impact

An attacker who shares a notebook or a Markdown file - via email, GitHub, or a Binder link - can invoke an arbitrary command upon a single click by the victim. The button can be rendered inside the output area and be visually indistinguishable from a legitimate widget. No kernel needs to start; the HTML output is stored in the notebook file and displayed immediately on open.

Single-click impact

An attacker convincing the victim to click on a single button or link can:

• execute arbitrary code in the available kernels,
• delete files leading to information loss; in principle the loss could be unrecoverable, depending on server configuration and attack complexity,
• open multiple kernels/terminals at once, or create multiple files at once, putting significant stress on the server and thus deny availability for other users when using standalone multi-tenant jupyter-server deployment, and to a lesser degree impact availability on JupyterHub deployments.

The arbitrary code execution will be immediately visible to the user; and can be halted by the timely user intervention. The deletion of files can be silent and go unnoticed for some time.

Multi-click attacks

An attacker who convinces the victim to click on multiple buttons in specific order and to grant access to clipboard (or in scenarios where the user already granted keyboard access) can obtain full access to the terminal and execute arbitrary commands in the environment with access scope that might exceed that of available kernels. Only users of Chromium-based browsers are susceptible to this expanded variant of the attack.

The execution of commands in the terminal would be immediately visible to the user.

Impact of third-party extensions

The impact described above assumes a plain JupyterLab/Notebook installation. In environments with frontend extensions that contribute additional commands the attack surface is increased by the functionality covered by these commands.

Patches

JupyterLab 4.5.7

Workarounds

No workarounds are available for end-users.

Downstream applications inheriting from JupyterFrontEnd or JupyterLab can effectively disable the CommandLinker by passing commandLinker: new CommandLinker({ commands: new CommandRegistry() }) option in the initialization options.

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-files

Impact

A stored Cross-Site Scripting (XSS) vulnerability in Jupyter Notebook allows attackers to steal authentication tokens from users who open malicious notebook files and interact with elements that the attacker can make look indistinguishable from legitimate controls (single click interaction).

The vulnerability enables complete account takeover through the Jupyter REST API, allowing the attacker to:

1. Read all files
2. Modify/create files
3. Access running kernels and execute arbitrary code
4. Create terminals for shell access

Patches

Jupyter Notebook 7.5.6 and JupyterLab 4.5.7 include patches for this vulnerability.

Workarounds

The help extension can be disabled via CLI:

jupyter labextension disable @jupyter-notebook/help-extension
jupyter labextension disable @jupyterlab/help-extension

Hardening

The patched versions include a toggle to disable the command linker functionality altogether, for example via overrides.json:

{
  "@jupyterlab/apputils-extension:sanitizer": {
    "allowCommandLinker": false
  }
}

Resources

• https://jupyterlab.readthedocs.io/en/latest/user/commands.html#commands-in-markdown-output-and-files

Acknowledgments

Reported by Daniel Teixeira - NVIDIA AI Red Team