GHSA-833v-v58w-294h (CVE-2026-76807, CVSS 9.3, critical) reached the GitHub Advisory Database on November 19, 2026, after NVD published the CVE on November 18. Versions of @doist/todoist-mcp up to and including 10.3.2 ran an optional HTTP server, todoist-mcp-http, that any website could drive through DNS rebinding. The server then executed Todoist tools with the operator’s own TODOIST_API_KEY. If you run that server, update to 10.3.3 now. If you only use the package over stdio, this advisory does not reach you.
This is a textbook case of a local MCP server trusting "it only listens on loopback." The same class of bug shows up whenever a coding agent's tools are exposed over HTTP on your own machine, so it is worth understanding even if you never touch Todoist.
TL;DR
| Question | Answer |
|---|---|
| Am I affected? | Only if you run todoist-mcp-http (the optional Streamable HTTP server) at version 10.3.2 or older. The stdio entrypoint is not browser-reachable. |
| Is there a fix? | Yes: @doist/todoist-mcp 10.3.3, per the advisory and the v10.3.3 release. |
| What can an attacker do? | Call any registered Todoist tool as you: create, update, complete and delete tasks, and change projects, comments, reminders, labels, filters and goals. |
| Does the attacker need my token? | No. The server holds the token; the browser supplies none. |
| Does data leave my machine? | Todoist data is returned to the attacker's page in the MCP responses, since the page is the caller. |
| Exploited in the wild? | The advisory describes a proof of concept and a demo video; it does not report in-the-wild use. |
What is CVE-2026-76807?
The advisory describes the self-hosted todoist-mcp-http server as accepting browser-origin MCP requests with attacker-controlled Host and Origin headers. The server is started with a command like TODOIST_API_KEY=your-key npx -p @doist/todoist-mcp todoist-mcp-http, and it binds to 127.0.0.1 on port 3000 by default (HOST and PORT can change that).
Per the advisory, the code path is short. POST /mcp hands the request to the SDK's StreamableHTTPServerTransport without checking Host or Origin. A middleware does run, but it only validates the server's Todoist token against Todoist and then calls next(). It never authenticates the caller. Every tool is then registered against the server-side API client: the advisory names add-tasks, complete-tasks, update-tasks, delete-object, comments, reminders, labels, filters, goals and assignments.
The reporter is credited as Avishai Gonen of Pluto Security. The affected package is the npm package @doist/todoist-mcp; the advisory lists <= 10.3.2 as vulnerable and 10.3.3 as the first patched version, with the fix in pull request 516.
How does DNS rebinding reach a loopback-only server?
Browsers protect local services with the same-origin policy, which ties a page to its hostname. DNS rebinding sidesteps that by controlling the DNS answer for the hostname.
- The victim visits an attacker page, say
evil.example, which resolves to the attacker's server first. The page loads and runs JavaScript. - The attacker's DNS server changes the answer for
evil.exampleto127.0.0.1with a very short TTL. - The browser re-resolves the name for the next request. It still considers the origin to be
evil.example, so same-origin rules are satisfied, but the TCP connection now goes to the victim's own machine, to the port the page chose (for example 3000). - The request arrives at
todoist-mcp-httpcarryingHost: evil.exampleandOrigin: https://evil.example. Because the server never checks them, it treats the request like any other. - The page sends MCP JSON-RPC to
/mcp:tools/listto enumerate tools, thentools/callwithuser-info,add-tasks,complete-tasksand so on. - The server executes each call with its own API key and returns the result to the page, which can read it because the origin matches.
The advisory's proof of concept followed exactly this path with a throwaway Todoist account. user-info returned the account tied to the server-held key. add-tasks created a task titled "DNS REBINDING DEMO - attacker-created task", and complete-tasks marked it done, confirmed in the Todoist UI. The reporter also linked a demo video. The demonstrated impact is task creation and completion; the advisory states that the same path exposes the rest of the write surface, including deletion.
No browser exploit, no malware and no stolen credential is needed, which is why the CVSS vector reads network-reachable, low complexity, no privileges, with user interaction (visiting the page) and a changed scope.
Why "I only bind to localhost" is not an answer
The advisory's own wording is useful: users "reasonably treat" a loopback bind as local-only, and DNS rebinding "preserves the attacker's Host and Origin while routing the TCP connection to 127.0.0.1." Binding to loopback stops other machines on your network from connecting. It does nothing to stop your own browser from connecting on behalf of a web page.
The MCP project documents this. The Streamable HTTP section of the MCP transports specification warns that without protection, DNS rebinding can be used to interact with local MCP servers from remote websites, and tells implementers to validate the Origin header on incoming connections, bind only to 127.0.0.1 when running locally, and implement proper authentication. The Todoist server did the second and skipped the other two.
What changed in 10.3.3
The advisory's remediation section describes what a fix should contain, and 10.3.3 is the patched release (see the fix commit for the exact change). The recommended controls are:
- Strict
Hostvalidation before MCP dispatch: accept only explicit loopback names such as127.0.0.1,localhostand[::1], plus any operator-configured trusted names. - Reject browser requests with a missing or untrusted
Originbefore they reach the transport. - A per-run bearer token or equivalent caller secret for HTTP mode, even on loopback, because Host and Origin checks reduce the exposure but are not caller authentication.
- Enable the MCP SDK's DNS rebinding protection with explicit allowed hosts and origins, and keep stdio as the default local transport.
I did not diff the 10.3.3 code line by line for this post, so read the release notes and commit before relying on a specific control. The affected and patched version numbers come from the advisory.
What this means if you run AI coding agents
1. Find out which transport you use. An MCP config entry with command and args launches the server as a child process over stdio; that is not browser-reachable. An entry with a url such as http://127.0.0.1:3000/mcp, or a terminal where you started todoist-mcp-http by hand, means you run the HTTP server.
2. Check the version and update. Typical commands: npm ls -g @doist/todoist-mcp for a global install, or re-run through npx with an explicit version such as npx -p @doist/[email protected] todoist-mcp-http. Confirm the version you end up with is 10.3.3 or later.
3. If you cannot update right away, stop the HTTP server or switch the agent to the stdio entrypoint. Do not leave it running while you browse.
4. Look for signs of use. Review Todoist's activity history for tasks, comments or deletions you did not make, especially around times you had the HTTP server running. The advisory describes no in-the-wild exploitation, so this is a precaution. Rotating the API token is cheap if anything looks off.
5. Audit your other local HTTP MCP servers the same way. Any server listening on 127.0.0.1 that does not validate Host/Origin and does not require a caller secret has the same exposure. A quick test: from a shell, send a request with a bogus header, for example curl -i -H "Origin: https://example.com" -H "Host: evil.example" http://127.0.0.1:3000/mcp, against a server you own. A server that answers instead of rejecting is a candidate for the same bug. Run this only on your own machine.
6. Prefer stdio for local tools. It removes the browser from the picture entirely, which is also the advisory's recommendation.
What Beam can and cannot do here
Beam is local, privacy-first security and observability for AI coding agents, and its documented scanner can help with the configuration part of this problem. beam scan .mcp.json --mcp runs offline heuristic checks on an MCP config, validates the JSON and examines version pinning, as described in the scanning docs. Running it on your config is a reasonable way to review which MCP entries you have and whether packages are pinned to a version you can check against an advisory.
The limits matter. The scan does not launch the server, does not contact it, and does not test whether a running HTTP server validates Host or Origin. It is not semantic analysis and it will not know that 10.3.2 is vulnerable. This is a flaw in a server's network handling, which Beam's hook-based deny on agent tool calls does not intercept: the malicious requests come from a web page to a local port, not from your agent. Update the package; use Beam to keep an inventory and notice unpinned MCP entries. For background, see our practical MCP security guide and how Beam keeps data on-device.
Context: local servers keep being the weak point
This is not an isolated pattern. In November, an AWS bulletin covered an SSRF in mcp-gateway-registry, where a gateway became an internal-network client. In October, Node.js shipped a security release affecting MCP servers and agent tools. The common thread is that MCP servers are small network services with powerful credentials, and their authors often borrow the trust model of "it is only local."
Tool servers that hold a long-lived API key and expose write operations deserve the same scrutiny as any web service: authenticate callers, validate headers, and keep the key's scope as narrow as the provider allows.
Frequently asked questions
What is CVE-2026-76807?
It is a critical (CVSS 9.3) flaw in the optional todoist-mcp-http server of @doist/todoist-mcp. The server accepted browser requests with any Host and Origin header and ran Todoist tools with the operator's API key, so a website using DNS rebinding could operate the local server.
Which versions of the Todoist MCP server are affected?
The advisory lists @doist/todoist-mcp up to and including 10.3.2, and the first patched version is 10.3.3. Only the optional Streamable HTTP server (todoist-mcp-http) is browser-reachable; the stdio entrypoint is not.
Is there a DNS rebinding protection setting for MCP servers?
The MCP specification tells local HTTP servers to validate the Origin header, bind to 127.0.0.1 only, and authenticate connections. The advisory also recommends strict Host validation and a per-run bearer token, and says the MCP SDK's DNS rebinding protection should be enabled with explicit allowed hosts and origins where available.
Do I need to rotate my Todoist API token?
The advisory does not report exploitation in the wild, and the demonstrated attack used the server, not a stolen token, so the token itself is not shown to leak. If you ran an old HTTP server while browsing untrusted sites and see tasks you did not create, review your Todoist activity and rotate the token as a precaution.
Does running the server on 127.0.0.1 keep me safe?
No. DNS rebinding makes a public website's origin resolve to 127.0.0.1, so the browser sends requests to your loopback server with the attacker's Host and Origin. Loopback binding only helps if the server also checks those headers or requires a caller secret.
How do I check whether I run the vulnerable HTTP server?
Run npm ls -g @doist/todoist-mcp or check the version in your npx cache, and look for a process started as todoist-mcp-http listening on port 3000 or your PORT value. If your MCP config launches the package over stdio, you are not running the browser-reachable component.
Related reading
- MCP security: a practical guide
- CVE-2026-77359: Jupyter MCP Server DNS rebinding on management routes
- CVE-2026-125926: AWS mcp-gateway-registry SSRF
- Node.js October 2026 security release and MCP servers
- Beam on your machine: what stays local
Primary sources: GitHub advisory GHSA-833v-v58w-294h, NVD record for CVE-2026-76807, MCP transports specification.
Versions, advisory details and fix status are accurate as of November 19, 2026, the day the advisory was added to the GitHub Advisory Database. Check the advisory and the package's release notes for later changes.

