CVE-2026-77359: Jupyter MCP Server's /api/connect and /api/stop routes skipped the token and Origin checks, so a website or network caller could repoint it to an attacker's Jupyter backend (fixed 1.0.3)
In standalone streamable-http mode, Datalayer's Jupyter MCP Server (pip package jupyter-mcp-server) protects /mcp with MCP_TOKEN and rejects forged Host and Origin headers, but up to 1.0.2 the custom management routes on the same app, PUT /api/connect and DELETE /api/stop, had neither check. An unauthenticated network caller (CVE-2026-77318), or a malicious website using DNS rebinding against a loopback, LAN, VPN or container-published listener (CVE-2026-77359), could replace the runtime and document URLs and tokens, reset the server context and make the server start a kernel client against attacker-controlled infrastructure, or stop the current notebook kernel. Fixed in 1.0.3.
What it is
jupyter_mcp_server/server.py registers /api/connect (PUT) and /api/stop (DELETE) with mcp.custom_route. /api/connect parses the caller's JSON, calls set_config with the supplied provider, runtime_url, runtime_token, document_url and document_token, resets ServerContext and starts the kernel.
CVE-2026-77318 (GHSA-24pw-rf5p-mgp7, high): MCP_TOKEN is enforced on /mcp but not on these routes, so any caller that can reach the port can use them without credentials. This happens even without the explicit --insecure-mcp-noauth mode.
CVE-2026-77359 (GHSA-9x8x-crjq-g64h, critical): the Host/Origin rejection that /mcp applies (HTTP 421 Invalid Host header) is not applied to the management routes. In the advisory's proof of concept a browser-shaped request with an attacker Origin gets 200 {"success":true}, and the attacker's backend then receives Jupyter API traffic from the MCP server, including POST /api/kernels.
Affected: jupyter-mcp-server <= 1.0.2 in standalone streamable-http mode. Version 1.0.3 was released on 11 July 2026; GitHub published both advisories and the CVEs on 19 November 2026.
How to check
- Run pip show jupyter-mcp-server and check whether you start it with --transport streamable-http.
- Check which address it listens on. A broad bind (0.0.0.0, LAN or a published Docker port) makes it reachable by other machines as well as by DNS rebinding.
How to fix
- Upgrade jupyter-mcp-server to 1.0.3 or later.
- Bind the HTTP transport to 127.0.0.1 unless you need remote access, and put remote access behind an authenticating proxy.
- If an affected server was reachable, rotate the Jupyter runtime and document tokens it was configured with.
pip show jupyter-mcp-server | grep -i '^version'
# which address is the MCP HTTP server bound to?
ss -ltnp | grep -i python