gitlab-mcp (zereight) Server-Side Request Forgery Leading to Authentication Token Theft

Release Date: 15th September 2026 (Added 16 September 2026)

Impact : CRITICAL

TLP Rating: Clear

The Department of Communications and Digital Transformation (DCDT) through CERT Vanuatu (CERTVU), provides the following advisory.

This alert is relevant to Organizations and System/Network administrators, and development teams, that run the gitlab-mcp server to connect AI coding assistants (such as Claude, VS Code, or Cursor) to a GitLab instance. This alert is intended to be understood by technical users and systems administrators.

What is it?

CVE-2026-61559 is a critical vulnerability in gitlab-mcp, a popular open-source Model Context Protocol (MCP) server (over 1,900 GitHub stars) that lets AI coding assistants manage GitLab projects, merge requests, issues, pipelines, and wikis on a user's behalf - the same general category of AI-agent connector tooling CERTVU previously judged relevant to Vanuatu in Advisory 237 (LiteLLM MCP Authentication Bypass). When the server is run with the ENABLE_DYNAMIC_API_URL configuration option set to true, it reads an attacker-controllable X-GitLab-API-URL request header without validating the hostname, and attaches the victim's own GitLab authentication token to API calls sent to that attacker-specified URL, allowing an attacker to steal a victim's GitLab token via server-side request forgery.

Reference

  1. https://www.cve.org/CVERecord?id=CVE-2026-61559
  2. https://github.com/zereight/gitlab-mcp/security/advisories/GHSA-2h44-8472-frjj