On this page (8 sections)
CVE-2026-77359 affects jupyter-mcp-server, the Model Context Protocol server for Jupyter, in versions up to and including 1.0.2. In standalone streamable-http mode, the server checks Host and Origin headers on its /mcp endpoint but does not apply the same browser-origin controls to two management routes, /api/connect and /api/stop. According to the CVE record, a malicious website can use DNS rebinding to reach a listener that was never meant to be exposed to web pages.
If you run this package on a laptop, a shared LAN host, behind a VPN or in a published container, check your version and upgrade to 1.0.3. The advisory is published on GitHub.
What happened
According to the GitHub advisory, the /mcp endpoint rejects forged Host and Origin headers. In the advisory's example, a request with a forged Host value gets an HTTP 421 Invalid Host header response. Custom management routes mounted on the same application do not get that protection.
Two of those routes matter. PUT /api/connect replaces the active Jupyter backend settings, including runtime_url and document_url, and starts a kernel client. DELETE /api/stop stops the current notebook kernel. The advisory's demonstration was run against /api/connect, which returned a success response to a browser-shaped request. For /api/stop, the advisory states that it is also state-changing and accepts the same browser-origin route shape.
How the attack works in plain terms
DNS rebinding is a way to make a victim's browser talk to a private address while still believing it is talking to the attacker's website. The attacker's hostname first resolves to the attacker's server, so the page loads and its script runs under that origin. The hostname is then re-pointed to a loopback or private address. The browser still sees the same hostname, so the script can send requests to the internal service.
The usual defense is for the service to check the Host and Origin headers and refuse unexpected names. The advisory says /mcp does this and the management routes do not. The CVE record names the reachable targets as a loopback, LAN, VPN or published-container MCP listener.
What an attacker can achieve
- Replace the runtime and document URLs through /api/connect, so the MCP server starts a kernel client against infrastructure the attacker controls.
- Cause Jupyter API and WebSocket traffic from the MCP server to reach that infrastructure. In the advisory's example, the attacker's listener received requests carrying the attacker-supplied token.
- Stop the current notebook kernel through /api/stop.
The CVE record notes that the default broad bind can increase exposure, but states that the vulnerability is the missing origin boundary on the management routes, not a lack of Host validation on /mcp.
Who is affected
You are in scope if both of the following are true:
- You installed the pip package jupyter-mcp-server at version 1.0.2 or earlier.
- You run it in standalone streamable-http mode, on a host whose address a browser on the same machine or network could reach (the advisory lists loopback, LAN, VPN and published-container listeners).
A token configured for /mcp did not prevent the demonstrated issue. The advisory's demonstration ran with MCP_TOKEN configured, and the /api/connect request was still accepted.
How to check your exposure
Run this inside the Python environment or container where the server is installed. It prints the installed version; 1.0.2 or lower is affected.
python -m pip show jupyter-mcp-serverTo see which address the server listens on, list listening sockets and look at the address next to the port you configured. A wildcard address such as 0.0.0.0 means other machines can reach it. The advisory lists loopback among the reachable targets, so a loopback bind should not be treated as a substitute for upgrading.
ss -ltnpWhat to do now
- Upgrade jupyter-mcp-server to 1.0.3 or later in every environment where it is installed, including any container images you maintain.
- Restart the server after upgrading and confirm the version again with pip show.
- As general hardening, bind the server to a specific trusted interface rather than a wildcard address where you can, and avoid publishing its port from containers unless needed.
- As general hardening, restrict network access with a firewall so only the clients that need the port can reach it.
- If you cannot upgrade yet, consider stopping the server when it is not in use.
python -m pip install --upgrade "jupyter-mcp-server>=1.0.3"Detection and monitoring
The advisory's server log shows a line beginning with Connect endpoint received, listing the runtime_url and document_url, followed by a kernel start message. If your logs are retained, search them for connect events whose URLs you do not recognize.
grep -n "Connect endpoint received" /path/to/your/server.logAs a precaution, if you find a connect event pointing at an unexpected destination, consider any runtime or document token used with this server to be at risk and rotate it. Reviewing outbound connections from the host for unfamiliar destinations is also reasonable general practice. These are suggestions, not findings from the advisory.
A network firewall or web application firewall is one layer of defense for services you expose, though it does not replace patching. For background, see What is a WAF? and our small business website security guide.
Timeline
| Date | Event |
|---|---|
| Jul 12, 2026 | GitHub advisory GHSA-9x8x-crjq-g64h shown as published |
| Nov 19, 2026 | CVE-2026-77359 published in the CVE record |
| Nov 20, 2026 | GitHub advisory last updated |
Sources
This article is based on the public sources below. Details can change as vendors and researchers publish more; check the original advisories before acting on version numbers.
