Two flaws let attackers hijack Jupyter MCP Server's backend
Advisories published on 20 November 2026 say unauthenticated management routes in Datalayer's Jupyter MCP Server can be redirected to attacker-controlled infrastructure.
By The Agentic Times ·

Two security advisories published on 20 November 2026 describe flaws in Datalayer's Jupyter MCP Server, a Python package that lets AI agents drive Jupyter notebooks. Both affect the standalone streamable-http mode, and both allow an attacker to change which Jupyter backend the server talks to without holding the server's authentication token.
The first, tracked as CVE-2026-77318, concerns missing authentication on custom HTTP management routes. According to the advisory, the server requires an MCP_TOKEN bearer token on the main /mcp endpoint, and its command-line interface refuses to start the HTTP transport without one unless the operator explicitly passes an --insecure-mcp-noauth flag. But the PUT /api/connect and DELETE /api/stop routes, mounted on the same application, do not check that token.
The advisory includes a proof of concept. With MCP_TOKEN set, an unauthenticated request to /mcp returns 401 Unauthorized. The same unauthenticated caller can then send a JSON body to /api/connect naming its own runtime_url, document_url and tokens, and receives {"success":true}. Server logs in the advisory show the MCP server resetting its context and starting a kernel client against the supplied address, after which the attacker's listener receives Jupyter API calls and WebSocket traffic carrying the attacker-chosen bearer token. The /api/stop route can likewise be used to remove the current notebook.
The second advisory, CVE-2026-77359, describes the same routes being reachable from a victim's web browser through DNS rebinding, a technique in which a malicious site keeps its own origin while its hostname resolves to a private address such as 127.0.0.1. The advisory shows that /mcp rejects forged Host and Origin headers with HTTP 421 Invalid Host header, even when a valid bearer token is supplied. The custom management routes do not apply that check. A PUT /api/connect request carrying Host: attacker.example:8083 and a matching Origin header was accepted and returned HTTP 200, with the server again starting a kernel client against the attacker's backend.
The advisories argue this is a genuine boundary failure rather than intended behaviour. Both note that the proof-of-concept runs are not using the explicit insecure mode, that the vulnerable path sits outside normal JSON-RPC tool dispatch so is not a trusted MCP-client operation, and that the issue is not simply the server's default 0.0.0.0 bind. That broad bind increases exposure, the advisories say, but the core problem is that state-changing routes bypass the controls applied to /mcp.
On exposure, the DNS rebinding advisory states that a web attacker could reach a local, LAN, VPN or Docker-published MCP listener from a victim's browser. The direct advisory frames its case as a same-network attacker making an unauthenticated state change against the HTTP listener.
For remediation, the second advisory recommends applying the same browser-origin policy to every route mounted on the standalone HTTP application. Specifically, it says Host and Origin validation should run before custom routes parse request bodies or mutate global configuration, and that the MCP_TOKEN bearer token should be required on /api/connect, /api/stop and the other custom routes. Both advisories reference the project's security documentation, which states that all MCP endpoints require bearer token authentication and that client authentication has been required for the streamable-http transport since version 1.0.0.
Sources
- github.comhttps://github.com/advisories/GHSA-24pw-rf5p-mgp7
- github.comhttps://github.com/advisories/GHSA-9x8x-crjq-g64h
