Two security flaws, one rated critical, in Jupyter MCP Server allow attackers to hijack the Jupyter backend by exploiting unprotected management routes in standalone streamable-http mode. The vulnerabilities, tracked as CVE-2026-77359 (CVSS 9.3) and CVE-2026-77318 (CVSS 7.3), affect all versions up to and including 1.0.2.
Quick Take
- What: Two vulnerabilities in Jupyter MCP Server let attackers replace the Jupyter backend URL and tokens through unauthenticated
/api/connectand/api/stoproutes - Who is affected: Users running jupyter-mcp-server versions up to 1.0.2 in standalone streamable-http mode, especially those with exposed network listeners
- Severity: Critical (CVSS 9.3 for CVE-2026-77359, High 7.3 for CVE-2026-77318)
- What to do: Upgrade jupyter-mcp-server to version 1.0.3 or later immediately
- Status: Patch available in version 1.0.3
In plain English
Jupyter MCP Server is a tool that connects AI agents to Jupyter notebooks. Two security holes let attackers take control of which Jupyter server the MCP server talks to. The main /mcp endpoint correctly checks passwords and blocks fake requests, but two other routes (/api/connect and /api/stop) do not. An attacker can send requests to these routes to point the server at their own fake Jupyter server, or stop the current notebook. If you use this tool, update to version 1.0.3 right away. If you don’t use Jupyter MCP Server, you can ignore this.
What happened
Jupyter MCP Server is a Model Context Protocol server that connects AI agents to Jupyter notebooks. The project is maintained by Datalayer and distributed via pip as jupyter-mcp-server.
In standalone streamable-http mode, the server enforces authentication and origin validation on the main /mcp endpoint. When MCP_TOKEN is configured, requests to /mcp without valid credentials receive a 401 Unauthorized response. The endpoint also rejects forged Host and Origin headers with HTTP 421 Invalid Host header to prevent browser-based attacks.
However, two custom management routes mounted on the same application do not apply these security controls:
PUT /api/connectaccepts JSON containingruntime_url,runtime_token,document_url, anddocument_tokenfields, updates the server’s active Jupyter backend configuration, resets the server context, and starts a kernel client against the specified infrastructure.DELETE /api/stopstops the current notebook kernel.
Both routes are defined in jupyter_mcp_server/server.py and process requests without checking the MCP_TOKEN bearer token or validating browser-origin headers.
CVE-2026-77359: DNS rebinding attack
CVE-2026-77359 (CVSS 9.3, Critical) describes a browser-based DNS rebinding attack. A malicious website can use DNS rebinding so that its hostname resolves to 127.0.0.1, a LAN address, a VPN address, or a published container port where an MCP listener is running. The attacker’s page keeps its own origin while the hostname resolves to the target address, then sends a PUT /api/connect request.
Because /api/connect does not validate Host and Origin headers, it accepts the browser-shaped request and replaces the MCP server’s active Jupyter backend with attacker-controlled infrastructure. The server then sends Jupyter API and WebSocket traffic to that infrastructure.
The GitHub advisory GHSA-9x8x-crjq-g64h states that the protected /mcp endpoint shows the intended browser-origin boundary, but custom management routes bypass it.
CVE-2026-77318: Unauthenticated network access
CVE-2026-77318 (CVSS 7.3, High) describes direct unauthenticated network access. Even when MCP_TOKEN is configured and the CLI enforces authentication for the /mcp endpoint, the /api/connect and /api/stop routes do not require the token.
An unauthenticated network caller can invoke PUT /api/connect to replace runtime_url and document_url values and their tokens, reset ServerContext, and start a kernel client against attacker-controlled infrastructure. The caller can also use DELETE /api/stop to stop the current notebook kernel.
The GitHub advisory GHSA-24pw-rf5p-mgp7 notes that this defeats the documented standalone HTTP security boundary because the protected /mcp endpoint correctly returns 401 Unauthorized without credentials, while adjacent management routes perform sensitive state changes without the same control.
Proof of concept
Both advisories include proof-of-concept demonstrations. Running commit 27b91e551d8bbca2da32c334f9f13267290ff815 (prior to 1.0.3) with MCP_TOKEN=SCAN_MCP_TOKEN, a request to POST /mcp without credentials returns 401 Unauthorized. A request with forged Host: attacker.example:8083 and Origin: attacker.example:8083 headers returns 421 Invalid Host header.
However, sending PUT /api/connect with the same forged headers or without credentials returns 200 OK with {"success: true}. The server log shows Connect endpoint received” and “Default notebook kernel started successfully”. When an attacker-controlled backend is listening on the specified address, it receives Jupyter API traffic with the attacker-chosen bearer token.
What you should do
Upgrade jupyter-mcp-server to version 1.0.3 or later. The NVD entries and GitHub advisories state that version 1.0.3 fixes both vulnerabilities.
If you cannot upgrade immediately, do not expose the streamable-http listener beyond localhost. Avoid the default 0.0.0.0 bind, and keep the listener off shared LANs, VPNs, and published container ports.
What it means
These vulnerabilities allow attackers to hijack the Jupyter backend connection of an MCP server, causing it to send API and WebSocket traffic to attacker-controlled infrastructure. The attacker can also stop the current notebook kernel.
The default broad bind (0.0.0.0) increases exposure, but the core issue is that state-changing custom routes bypass the authentication and browser-origin controls applied to the main /mcp endpoint. This occurs outside normal MCP JSON-RPC tool dispatch and does not require the explicit --insecure-mcp-noauth mode.
Only standalone streamable-http mode is affected. The sources do not report exploitation in the wild.
Exposed MCP management endpoints and DNS rebinding keep turning up in AI agent tooling: compare our reports on the pgEdge Postgres MCP Server DNS rebinding flaw and the Todoist MCP server rebinding bug. New cases are tracked in our AI Security hub.
Sources
- NVD CVE-2026-77359
- GitHub Advisory GHSA-9x8x-crjq-g64h
- GitHub Advisory GHSA-24pw-rf5p-mgp7



