Skip to main content
Menu
Advisories

Summary The fix for CVE-2026-77359/CVE-2026-77318 (GHSA-9x8x-crjq-g64h) added a middleware that checks whether request.url.hostname is localhost/127.0.0.1/::1 before allowing access to the…

GHSA-F8RF-RCRX-H6R9Published Updated GitHub

Description

The fix for CVE-2026-77359/CVE-2026-77318 (GHSA-9x8x-crjq-g64h) added a middleware that checks whether request.url.hostname is localhost/127.0.0.1/::1 before allowing access to the /api/connect, /api/stop, and /api/healthz management routes. That check reads Starlette's parsed URL, which is built from the client-supplied Host header whenever one is present - not from the actual TCP peer address. A single request with a forged Host: localhost header, regardless of where it actually originates, passes the check.

Details

  • jupyter_mcp_server/server.py:105-169 - ManagementRouteSecurityMiddleware.dispatch and _is_local_hostname (lines 110-111) do a pure string comparison against the parsed Host header, never referencing request.client or scope["client"] (confirmed via grep - neither appears anywhere in this file).
  • Confirmed against the actual pinned dependency: starlette==1.6.0's datastructures.py builds URL.netloc from scope["headers"]'s host entry (validated only against a syntax regex) whenever present, falling back to scope["server"] (the real bind address) only if the header is absent or malformed.
  • utils.py:429 - the server still binds 0.0.0.0 by default, so this is reachable from the network, not just localhost.

Proof of Concept

Started the real server (--insecure-mcp-noauth --transport streamable-http).

curl -X PUT /api/connect -H "Host: evil.com" correctly returns 421 Misdirected Request.

The same request with -H "Host: localhost" and a full attacker-controlled payload (code_sandbox_url, document_url, tokens) is accepted - server logs confirm the connect() handler ran, DocumentCodeSandbox(**data) validated, and set_config() overwrote the live server's code_sandbox_url with the attacker's URL, followed by the server actually attempting to connect a kernel to attacker infrastructure.

Scoping

In --mcp-token (secure) mode, /api/connect and /api/stop still correctly require a valid bearer token after the Host check passes, so the spoof alone doesn't achieve full config hijack there - only /api/healthz is reachable via the spoof in that mode, which is a lower-severity status/info-disclosure gap, not a full takeover. The full critical-severity impact (unauthenticated backend hijack) is specific to --insecure-mcp-noauth mode, which your own CLI already labels "NOT recommended for production" - but that mode's entire security guarantee currently rests on this one check, and it fails completely.

Duplicate check

Read both existing advisories in full - neither describes Host-header string forgery as a bypass mechanism (GHSA-9x8x's PoC used DNS rebinding; this needs no DNS trickery at all, just one header). No commit since the fix (c2e98fc) has touched _is_local_hostname or the middleware. The existing regression test (tests/test_tools.py::test_management_routes_reject_forged_host) only asserts that a non-local-looking Host is rejected - there's no test for a forged Host: localhost from a genuinely non-local source, which is exactly the gap here.

Suggested Fix

Check the actual ASGI scope["client"] peer address (or require the literal TCP connection to be 127.0.0.1) instead of, or in addition to, the client-supplied Host header.

I'd appreciate being credited by name (Syed Anas Mohiuddin) if this results in a fix, changelog entry, or advisory - happy to provide any additional detail needed for that.

Reported by: Syed Anas Mohiuddin, independent security researcher

Affected versions

References

GitHub · datalayer/jupyter-mcp-server