CVE-2026-76807: DNS rebinding in Todoist MCP server exposes accounts to malicious websites

A critical DNS rebinding vulnerability in the Todoist MCP server's optional HTTP transport allowed malicious websites to read and modify Todoist accounts by reaching the loopback listener without authentication. The flaw affected version 10.3.2 and earlier when running todoist-mcp-http. Version 10.3.3 patches the issue.

Christopher Ort
By Christopher Ort · 2 min read
Share
A developer's laptop screen showing a terminal window with a local web server running, while a browser window in the background displays a malicious website attempting to connect to localhost
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. What you should do
  5. What it means
  6. Sources
i10X Quick Take

Quick Take

  1. What: CVE-2026-76807, a DNS rebinding vulnerability in the Todoist MCP server’s optional HTTP transport (todoist-mcp-http), allowed malicious websites to invoke MCP tools using the server’s Todoist API key without authenticating the browser caller.
  2. Who is affected: Users running todoist-mcp-http version 10.3.2 or earlier. The stdio transport is not affected.
  3. CVSS score: 9.3 (Critical) in the GitHub advisory, which NVD lists; a proof of concept is public
  4. Fixed version: 10.3.3
  5. What to do: Upgrade the npm package @doist/todoist-mcp to version 10.3.3 or later.

In plain English

The Todoist MCP server is a tool that lets AI agents read and change tasks in a Todoist account. It has an optional web server mode called todoist-mcp-http that runs on your computer. Before version 10.3.3, this web server did not check whether requests came from a trusted source. A malicious website could use a trick called DNS rebinding to send commands to the server running on your own machine, even though the server was only listening on localhost.

What you should do: If you use the Todoist MCP server in HTTP mode, you should upgrade to version 10.3.3.

What happened

The Todoist MCP server provides tools that allow AI agents to access and modify a Todoist account. The optional todoist-mcp-http server, which runs on loopback (127.0.0.1) by default, accepted browser-origin requests to POST /mcp without validating the Host and Origin headers or authenticating the browser caller.

A malicious website could use DNS rebinding to route its own origin to a victim’s local 127.0.0.1 MCP server and invoke MCP tools through StreamableHTTPServerTransport. The server used the operator’s TODOIST_API_KEY to execute these tools, exposing Todoist account data and permitting operations such as task creation, updates, completion, and deletion.

The vulnerability was present in version 10.3.2 and earlier of the npm package @doist/todoist-mcp. The stdio transport, which is not browser-reachable, was not affected by this attack path.

Avishai Gonen of Pluto Security reported the issue and provided a proof of concept demonstrating how an attacker could enumerate tools, retrieve user information, create tasks, and complete tasks through DNS rebinding.

What you should do

Upgrade the npm package @doist/todoist-mcp to version 10.3.3 or later. If you run todoist-mcp-http, make sure the running server is the updated version. Users who only use the stdio transport do not need to take action, as that mode is not affected.

What it means

This vulnerability violated the browser-origin boundary. Even though the MCP server ran only on loopback, DNS rebinding allowed a public website to reach it. The server validated its own Todoist backend token but did not authenticate the browser caller or validate Host and Origin headers.

The advisory’s remediation advice is to validate the Host header against explicit loopback names, reject browser requests with a missing or untrusted Origin, and require a per-run local bearer token for HTTP mode, because Host and Origin checks alone are not caller authentication. Version 10.3.3 is the patched release. For developers building MCP servers, the lesson is that binding to 127.0.0.1 is not a security boundary against a browser. Teams that let staff run local AI tooling should prefer the stdio transport, which the advisory calls the safest default.

Local MCP servers that listen over HTTP keep running into the same problem: see our reports on the pgEdge Postgres MCP Server DNS rebinding flaw and on the Dynatrace MCP Server allowing unauthenticated tool calls in HTTP mode. New cases are tracked in our AI Security hub.

Sources

  • NVD CVE-2026-76807
  • GitHub Security Advisory GHSA-833v-v58w-294h
Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Viable Attack Plans After Safeguards Were Bypassed
  2. ResolveAI SRE-1: A 27B Model Trained to Find the Root Cause of Production Incidents
  3. Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages
  4. Pika Enact Lets Creators Direct AI Video Actors and Cameras in Real Time
  5. Google Stitch Loop: An Always-On Product Agent That Fixes Security, Speed and Accessibility via Pull Requests