CVE-2026-102831: JupyterLab: clipboard cell-paste trust bypass enables XSS

GHSA-6966-vjj6-99xvHIGH
Published September 29, 2026
CISO Take

JupyterLab, Notebook and JupyterLite let a pasted code cell carry attacker-controlled metadata.trusted, so HTML output is treated as trusted and its script runs in the authenticated Jupyter origin without the cell ever being executed. The result is stored/DOM-style XSS with high confidentiality and integrity impact (CVSS 8.1). It needs user interaction, and the system clipboard paste feature must be active with pasteCodeCellsWithoutOutput disabled. The blast radius is wide: 1,892 downstream dependents, and Jupyter servers commonly hold cloud credentials, API keys, datasets and model artifacts. Exploitation risk today is moderate rather than urgent. EPSS is 0.00196 (top 91% percentile), it is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK. Patch to JupyterLab 4.5.11 or 4.6.4, Notebook 7.6.3 and JupyterLite Core 0.8.4. Until then, tell data scientists not to paste cells from untrusted sources, and prioritize multi-user and internet-reachable Jupyter deployments.

Sources:NVDEPSSGitHub AdvisoryATLASOpenSSFCISA SSVC

What is the risk?

Rated high (CVSS 8.1) but contextual risk is medium. Exploitation needs a victim to paste crafted cell JSON from the clipboard, plus a non-default-looking configuration (useSystemClipboardForCells active, pasteCodeCellsWithoutOutput disabled). EPSS is low (0.00196, top 91%), there is no KEV listing, no public PoC and no scanner template. Impact is high: script runs in the authenticated Jupyter origin, which can reach the kernel API, so it can lead to code execution as the user, token theft and data exfiltration. Exposure is greatest on shared JupyterHub, cloud-hosted notebooks and JupyterLite deployments where users routinely copy cells from tutorials, forums or chat. The package has an OpenSSF Scorecard of 5.8/10 and 94 other CVEs, which signals a recurring patch burden.

How does the attack unfold?

Payload Delivery
Attacker distributes crafted cell JSON with metadata.trusted=true and a malicious HTML output via a tutorial, forum, chat or ticket.
AML.T0011
User Paste
Victim pastes the cell using system clipboard paste, and the trusted flag is kept instead of being cleared.
AML.T0011
Sanitization Bypass
The HTML output renders as trusted and its script runs in the authenticated JupyterLab origin without executing the cell.
AML.T0049
Impact
The script abuses Jupyter APIs to steal tokens, API keys and datasets or to run code in the kernel.

What systems are affected?

PackageEcosystemVulnerable RangePatched
Jupyterpip>= 4.6.0, <= 4.6.34.6.4
13.5KOpenSSF 5.71.9K dependentsPushed 7d ago54% patched~27d to patchFull package profile →
Jupyterpip>= 0.7.0, <= 0.8.30.8.4
13.5KOpenSSF 5.71.9K dependentsPushed 7d ago54% patched~27d to patchFull package profile →
Jupyterpip>= 0.7.0, <= 0.8.30.8.4
13.5KOpenSSF 5.71.9K dependentsPushed 7d ago54% patched~27d to patchFull package profile →
Jupyter Notebookpip>= 7.5.0, <= 7.6.27.6.3
13.5KOpenSSF 5.73.0K dependentsPushed 7d ago80% patched~74d to patchFull package profile →

How severe is it?

CVSS 3.1
8.1 / 10
EPSS
0.2%
chance of exploitation in 30 days
Higher than 15% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Moderate

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRNone
UIRequired
SUnchanged
CHigh
IHigh
ANone

What should I do?

1 step
  1. 1) Upgrade JupyterLab to 4.5.11 or 4.6.4, Notebook to 7.6.3, and jupyterlite-core and jupyterlite to 0.8.4, and rebuild any static JupyterLite sites. 2) Workaround: disable useSystemClipboardForCells, or enable pasteCodeCellsWithoutOutput so pasted cells drop outputs and trust metadata. 3) Train users not to paste cells from untrusted sources and to review them in a plain-text editor first. 4) Harden deployments: set a strict CSP, isolate user servers on separate origins, use short-lived tokens, and avoid storing long-lived cloud or LLM API keys in notebook environments. 5) Detection: inventory versions with pip list or SBOM tooling, and audit notebooks for metadata.trusted:true in code cells with text/html outputs. Review Jupyter server logs for unexpected kernel or API calls shortly after paste events.

What does CISA's SSVC say?

DecisionTrack
Exploitationnone
AutomatableNo
Technical Impacttotal

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Code ExecutionData LeakageSocial EngineeringFrameworkAML.T0011AML.T0049

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Art. 15-Accuracy, robustness and cybersecurity
ISO 42001
A.6.2.6-AI system operation and monitoring
NIST AI RMF
MANAGE 2.3-Responses to previously unknown risks and mechanisms for supersession
OWASP LLM Top 10
LLM03-Supply Chain

Frequently Asked Questions

What is CVE-2026-102831?

JupyterLab, Notebook and JupyterLite let a pasted code cell carry attacker-controlled metadata.trusted, so HTML output is treated as trusted and its script runs in the authenticated Jupyter origin without the cell ever being executed. The result is stored/DOM-style XSS with high confidentiality and integrity impact (CVSS 8.1). It needs user interaction, and the system clipboard paste feature must be active with pasteCodeCellsWithoutOutput disabled. The blast radius is wide: 1,892 downstream dependents, and Jupyter servers commonly hold cloud credentials, API keys, datasets and model artifacts. Exploitation risk today is moderate rather than urgent. EPSS is 0.00196 (top 91% percentile), it is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK. Patch to JupyterLab 4.5.11 or 4.6.4, Notebook 7.6.3 and JupyterLite Core 0.8.4. Until then, tell data scientists not to paste cells from untrusted sources, and prioritize multi-user and internet-reachable Jupyter deployments.

Is CVE-2026-102831 actively exploited?

No confirmed active exploitation of CVE-2026-102831 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-102831?

1) Upgrade JupyterLab to 4.5.11 or 4.6.4, Notebook to 7.6.3, and jupyterlite-core and jupyterlite to 0.8.4, and rebuild any static JupyterLite sites. 2) Workaround: disable useSystemClipboardForCells, or enable pasteCodeCellsWithoutOutput so pasted cells drop outputs and trust metadata. 3) Train users not to paste cells from untrusted sources and to review them in a plain-text editor first. 4) Harden deployments: set a strict CSP, isolate user servers on separate origins, use short-lived tokens, and avoid storing long-lived cloud or LLM API keys in notebook environments. 5) Detection: inventory versions with pip list or SBOM tooling, and audit notebooks for metadata.trusted:true in code cells with text/html outputs. Review Jupyter server logs for unexpected kernel or API calls shortly after paste events.

What systems are affected by CVE-2026-102831?

This vulnerability affects the following AI/ML architecture patterns: Interactive ML development notebooks, JupyterHub multi-user environments, Browser-based JupyterLite/Pyodide notebooks, Training pipelines launched from notebooks, Managed notebook platforms.

What is the CVSS score for CVE-2026-102831?

CVE-2026-102831 has a CVSS v3.1 base score of 8.1 (HIGH). The EPSS exploitation probability is 0.24%.

What is the AI security impact?

Affected AI Architectures

Interactive ML development notebooksJupyterHub multi-user environmentsBrowser-based JupyterLite/Pyodide notebooksTraining pipelines launched from notebooksManaged notebook platforms

MITRE ATLAS Techniques

AML.T0011User Execution
AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Art. 15
ISO 42001: A.6.2.6
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

JupyterLab is an extensible environment for interactive and reproducible computing, based on the Jupyter Notebook Architecture. From JupyterLab 4.5.0 until 4.5.11 and 4.6.4, from Notebook 7.5.0 until 7.6.3, and from JupyterLite Core 0.7.0 until 0.8.4, the system clipboard cell-paste path accepts attacker-controlled cell JSON without clearing metadata.trusted. When useSystemClipboardForCells is active and pasteCodeCellsWithoutOutput is disabled, a pasted code cell can mark HTML output as trusted, bypass output sanitization, and execute script in the authenticated JupyterLab origin without executing the cell. Markdown and raw cells are not affected because their output is sanitized. This issue is fixed in JupyterLab 4.5.11 and 4.6.4, Notebook 7.6.3, and JupyterLite Core 0.8.4.

Exploitation Scenario

An attacker publishes a helpful-looking notebook snippet in a tutorial, forum post, chat or ticket, delivered as cell JSON that the browser copies to the clipboard. The cell has metadata.trusted set to true and a text/html output containing a script payload. A data scientist with the vulnerable JupyterLab pastes the cell into their notebook. The HTML output renders without sanitization and the script runs in the authenticated Jupyter origin without the cell being executed. The script then uses the Jupyter REST and kernel APIs to read environment variables and files, exfiltrate cloud and LLM API keys and datasets, or run code in the kernel to establish persistence.

CVSS Vector

CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:N

Timeline

Published
September 29, 2026
Last Modified
October 2, 2026
First Seen
September 29, 2026

Related Vulnerabilities