All posts

Nuxt Content’s query API flaw: can one request freeze your server?

CVE-2026-102650 (CVSS 7.5): the default @nuxt/content 3.x query endpoint can be made to stall a Node server with one request. Fixed in 3.15.2; use 3.16.2 or later.

Published
Reading time
6 min
By
Girish Babu GV

What happened#

Nuxt Content is the file-based content module for Nuxt. In version 3 it stores content in a local SQLite database and serves queries to the browser through an endpoint, POST /__nuxt_content/:collection/query. Per the advisory, that endpoint takes a raw SQL string from the request body and runs it. The only check is a denylist of a few keywords, which does not cover SQLite’s built-in functions that can allocate large blobs or do heavy work.

The advisory says a crafted request can force large memory allocations and heavy computation, and because the database library runs synchronously, that blocks the Node.js event loop and stalls all other requests. The reporters’ numbers are a baseline query of about 9 ms against roughly 12 seconds for a crafted one that was still accepted with an HTTP 200. The endpoint is enabled by default for every collection and needs no authentication.

The timeline on the advisory page: published by the maintainers on GitHub on 21 October 2026, by the NVD on 6 November, and added to GitHub’s advisory database on 10 November. The npm registry shows 3.15.2 was released on 23 July 2026, so the first fix has been available for a while, and the advisory itself came later.

Am I affected?#

Affected and fixed versions, as listed on the advisory
ItemWhat the advisory says
Package@nuxt/content (npm)
Affected3.0.0 up to, not including, 3.15.2
First patch3.15.2 (rejects function calls in query WHERE clauses)
Later hardening3.16.1 (validator ReDoS fix), 3.16.2 (only accepts WHERE clauses matching the query builder’s output)
Recommended3.16.2 or later
  • You use Nuxt Content 3.x with a server (server-rendered, or any deployment that keeps the query route): you are in scope if your version is below 3.15.2. The advisory names the default SQLite/D1 adapter as affected.
  • You prerender the whole site and deploy only static files: the advisory says the endpoint is not deployed, so you are not affected.
  • You are on Nuxt Content 2.x: the advisory’s affected range starts at 3.0.0, and I did not read anything about 2.x, so I can’t say more.

What to do now#

  1. Find your version in each project, including monorepo packages.
  2. Upgrade @nuxt/content to 3.16.2 or later, then test your content queries, since the later releases validate query input more strictly.
  3. If you can’t upgrade yet, the advisory lists two workarounds: block or restrict POST /__nuxt_content/*/query at a proxy or WAF, or prerender the site so the endpoint isn’t deployed. Both depend on your hosting, so check that your site still works afterwards.
  4. Check the route from outside. After upgrading or blocking, confirm from the public internet what the route does for you.
Check your version, then upgrade
# Which @nuxt/content versions are installed (direct and indirect)?
npm ls @nuxt/content

# Upgrade
npm install @nuxt/content@latest

How I checked this site#

This site is a Next.js project on Vercel. I searched its package.json for nuxt and found nothing, and in package-lock.json the only matches are peer-dependency declarations of other packages, not an installed nuxt or @nuxt/content package, so it does not use Nuxt Content. That covers this repository only.

What the sources don’t say#

  • Whether anyone has exploited it. The advisory does not claim exploitation. It shows an EPSS score of 0.35%, which is a probability estimate, not evidence.
  • How many deployments keep the query route reachable in production. That depends on each app’s hosting and rendering mode.
  • Whether the same endpoint is affected in other ways. The advisory also mentions two related reports from other researchers (GHSA-wjh8-qcrm-wf29 and GHSA-c739-vwm9-rgx3); I did not read those pages.
  • I did not read independent coverage. My second source is the npm registry, for versions and release dates.

My take#

My own opinion, not from the advisory.

The detail that stays with me is the pattern: an endpoint that accepts a query language from the browser and guards it with a denylist. A denylist for what an attacker may not run keeps losing to the next function nobody listed, and the later releases moving to “only accept what the query builder produces” is the sturdier design. If your own app exposes any endpoint that takes expressions rather than values, that is worth a second look.

Sources#

Keep reading

Esc