Security· Security Disclosure

Vite's November 2026 Dev-Server RCE: GHSA-p4f5-76q4-ffq7

Vite 8.3.7, 8.2.6, 8.1.7, 7.3.8, and 6.4.6 fix a critical RCE in the dev server. Only servers exposed with --host are hit. Patch, and keep dev on localhost.

Atlas Security
Security Advisories
15 min read · 3,234 words
Trevor Fannin@trevorfanninFounder, AtlasSecurity alerts and CVE breakdowns, plus infrastructure and general tech.Followon X (opens in a new tab)

Overview

Vite published GHSA-p4f5-76q4-ffq7 (opens in a new tab), "Remote code execution using __vite-browser-external virtual module," which it rates Critical, CVSS 4.0 9.1. The fixed releases — 8.3.7, 8.2.6, 8.1.7, 7.3.8, and 6.4.6 — went to npm on the morning of November 9, 2026, and the advisory followed that evening. Vite (pronounced "veet," the French word for "fast") is the build tool and dev server under most of today's front-end frameworks. The flaw is in its dev server: a crafted HTTP request can read files outside the project and run code on the machine running the server. The advisory's own summary: "By sending an HTTP request to the dev server, a arbitrary code execution was possible." There is one condition, and it is the whole story — only a dev server explicitly exposed to the network with --host or the server.host option is reachable. A build output, a preview server, or a dev server left on localhost is not affected.

Why it's urgent: the bug needs nothing but a reachable dev server, and that exact population is already under automated attack — for the weaker bugs that came before it. F5 Labs counted 807 session-grouped attacks, about 32,000 raw events, in August 2026 against internet-exposed Vite dev servers, harvesting .env files, AWS and Azure credentials, and Terraform state through an earlier file-read flaw (CVE-2026-39364). SANS Internet Storm Center logged the same class of scan in its honeypots in April 2026, and one of those file-read bugs, CVE-2025-31125, sits in CISA's Known Exploited Vulnerabilities catalog. This advisory turns that same exposure from reading a file into running code on the developer's machine. The patched releases shipped on the morning of November 9, hours before the advisory and nine days before Vite announced it on Bluesky. npm audit doesn't flag the bug yet (below), and a proof of concept is in the advisory. See the timeline.

Key facts

FieldDetail
AdvisoryGHSA-p4f5-76q4-ffq7 (opens in a new tab), "Remote code execution using __vite-browser-external virtual module." Vite's own repository advisory, published by maintainer sapphi-red
SeverityCritical, CVSS 4.0 9.1 (CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:H/VI:H/VA:N/SC:N/SI:N/SA:N). The rating is Vite's; there is no CVSS 3.x score
WeaknessCWE-116, "Improper Encoding or Escaping of Output"
ImpactArbitrary code execution on the host running the dev server, and read of files outside the allow list — only when the dev server is exposed to the network
Fixed invite 8.3.7, 8.2.6, 8.1.7, 7.3.8, 6.4.6 and Vite+ 1.5.1, published to npm on November 9, 2026. The advisory and Vite's announcement name 8.2.5 for the 8.2 line, but 8.2.5 was released on GitHub only, never on npm. The latest release is 8.3.8
No official fixVite 5.x and older: unsupported. The advisory's fix floor for everything below 6.4.6 is 6.4.6 itself (End-of-life versions)
ExploitationThe advisory includes a proof of concept. Not listed in CISA's KEV as of 11:30 AM CST, November 18, 2026
CVE recordsNone. No CVE ID, and nothing in NVD or GitHub's reviewed advisory database as of 11:30 AM CST, November 18, 2026. GitHub, a CVE Numbering Authority, has issued CVEs for Vite's past advisories, usually days later
Creditg2hsec, spansh, martinpriest, and zooyeon reported it, per the advisory

Who is affected

An application is affected only if two things are both true: it runs a vulnerable Vite version, and its dev server is reachable over the network. The version half:

Release lineStatus on Nov 18, 2026Fixed inNote
8.3 (current)Affected below 8.3.78.3.78.3.8 is the latest
8.2Affected at 8.2.4 and below8.2.6the advisory's 8.2.5 never reached npm
8.0 – 8.1Affected below 8.1.78.1.7one fix for both lines
7.xAffected below 7.3.87.3.8
4.0 – 6.4Affected below 6.4.66.4.6only the 6.4 line got a release
Vite+Affected at 1.5.0 and below1.5.1bundles its own Vite

The advisory lists 4.0.0 as the oldest affected version. Vite only ships security fixes for its current line and a few behind it, so the fix reached 8.3, 8.2, 8.1, 7.3, and 6.4 — the supported lines — and nothing older.

Vite+ and other builds outside the table

  • Vite+ (the vite-plus package, voidzero's unified toolchain) bundles its own copy of Vite and Rolldown, so its own version is the only one that matters. Vite+ 1.5.1 carries the fix; it bumped the bundled Vite from 8.3.5 to 8.3.7 and Rolldown from 1.2.16 to 1.2.18. The current release is 1.7.0. Inside a Vite+ project, the vite entry in your lockfile is an alias for @voidzero-dev/vite-plus-core, so check the Vite+ version, not that alias (below).
  • Frameworks that run Vite's dev server — Nuxt, SvelteKit, Storybook, Astro, and others — embed it rather than shipping their own. Whether this bug is reachable through one depends on the Vite version it resolves and on whether the framework's own server is exposed. We cover how to check them below.
  • A production build is not affected. The vulnerable code runs only in the dev server's request handling. vite build output, and the static files you deploy, don't contain it.

How the vulnerability works

When Vite's dev server can't bundle a module for the browser — a Node.js built-in like fs, or an optional peer dependency that isn't installed — it doesn't fail. It generates a tiny stand-in module in memory so your code still loads: Vite calls these virtual modules, and this bug is in two of them, __vite-browser-external and __vite-optional-peer-dep. The stand-in throws a helpful error if your code actually tries to use the missing module, and that error message includes the module's name.

The name came from the request URL, and Vite built the stand-in's JavaScript by dropping that name straight into the source without escaping it, so a crafted request could inject its own code into the stand-in. The dev server trusts the code it generates, so it treated the injected code as part of the project: the code could pull in files outside the server.fs allow list and make them readable, and it could run on the server itself. The result is file read, and then code execution, on the machine hosting the dev server, with the privileges of the user who started it.

None of this is reachable on a dev server bound to localhost, because the attacker can't send it a request. Exposing the server to the network with --host or server.host is what puts the request in reach — which is why that single condition decides whether you're affected.

Whose bug it is, and the fix

On Vite 8, the resolver that builds these stand-in modules lives in Rolldown, the Rust bundler that Vite 8 adopted — so the vulnerable code, and the fix, are in Rolldown, and Vite 8 picks it up through a dependency bump. The advisory notes that "the vulnerable code in v8 lives in the Rolldown binary," and because that code is only ever used by Vite, the report stays with Vite rather than Rolldown. On Vite 6 and 7, which don't use Rolldown for this, the code and the fix are in Vite's own source.

The fix is the same idea in both places: escape the module name before it goes into the generated JavaScript, so it can only ever be a string and can't break out into code. In Vite's source that's a JSON.stringify; in Rolldown it's an equivalent string-literal encoder. The stand-in now reads the name from a constant set to the escaped value.

Escape the module ID in the resolver's virtual modules
vitejs/vite · rolldown/rolldown · Nov 9, 2026
vite · resolve.ts, esbuildDepPlugin.tse5f23b6
Vite source · the 7.3.8 fix; 6.4.6 is the same change
17 lines added, 9 removed
rolldown · vite_resolve_plugin.rs27ac510
The v8 fix · ships in Rolldown 1.1.7, 1.2.18, and 1.3.1
14 lines added, 4 removed

Vite 8.3.7 and 8.2.6 require the fixed Rolldown, ~1.2.18 (the advisory names 1.2.17, which went to crates.io only); Vite 8.1.7 requires ~1.1.7. You get the Vite 8 fix by updating Vite, which moves its Rolldown with it — not by updating Rolldown yourself. Don't force a Rolldown version with overrides or resolutions either: an old one keeps the bug (below), and Rolldown's 1.3.0 release notes warn that Vite 8.3.6 and earlier need changes to work with Rolldown 1.3.

Which dev servers are exposed

The vendor's condition is narrow and precise: the dev server has to be reachable over the network. What follows is our reading of how that condition shows up in real setups — exposure, not a confirmed exploit path for any one of them.

SetupReachable from the network?This bug
vite / vite dev, defaultNo — binds to localhostNot reachable
vite --host, or server.host: true / '0.0.0.0'Yes — LAN, and public if the host isAffected
In Docker with --host and a published portYes, often publicAffected — see below
Framework dev server (Nuxt, SvelteKit, Storybook) exposedDepends on the framework's bind and Vite versionAffected if it runs a vulnerable Vite and is reachable
vite preview, or a production buildNo dev server involvedNot affected

Two things make exposure more common than it looks. Docker is the first: a container only reaches the outside through a published port, and getting Vite to answer from inside one means binding it to 0.0.0.0 with --host. Docker then publishes that port to every interface by default — its own docs call publishing "insecure by default" — and on Ubuntu hosts it writes its forwarding rules below ufw, so a port you published can be reachable even when your firewall says it shouldn't be. The second is remote and cloud dev environments, where --host is often the only way to reach the editor's preview at all, and the box already has a public address.

Patch the exposed servers first — anything started with --host, set to bind a non-loopback address, or published out of a container onto an untrusted network. A dev server that has only ever listened on localhost was never reachable, but patch it too, because the next --host is one command away.

How to check your systems

Run these in the project's repository, and on any host or image that runs the dev server. We tested the version and exposure checks with GNU grep 3.8, the current releases of npm, pnpm, Yarn, and Bun, and Vite 8.3.6 through 8.3.8 fixtures.

Check the Vite version

Your package.json shows a range like ^8.3.0; the lockfile shows what actually resolved. Read the lockfile. The quickest route is your package manager:

Shell
npm ls vite rolldown --all        # npm — --all walks nested copiespnpm list vite rolldown --depth Infinityyarn why vite                     # Yarn (classic and Berry)bun why vite                      # Bun

If you have no working install, read the lockfile directly:

Shell
# npm / Yarn / Bun: the resolved version sits under the package's blockgrep -nE -A1 '"node_modules/(.*/)?vite":|^"?vite@|"vite": \[' \  package-lock.json yarn.lock bun.lock 2>/dev/null# pnpmgrep -nE "^ +(vite|rolldown)@[0-9]" pnpm-lock.yaml

Four traps, each of which we reproduced:

  • On Vite 8, check Rolldown too. A fixed vite with an old rolldown pinned under it through overrides is still vulnerable. npm ls prints [email protected] overridden when that's the case — the fixed floors are 1.2.18 (for Vite 8.2 and 8.3), 1.1.7 (for 8.0 and 8.1), and 1.3.1 on the current Rolldown line.
  • A nested copy can be older than your top-level one. A project can resolve [email protected] at the top and a second, older vite under a tool like Vitest. npm ls vite --all (and --depth Infinity on pnpm) surface the second copy; a top-level-only check misses it.
  • Vite+ reports as vite. In a Vite+ project the vite entry is an alias for @voidzero-dev/vite-plus-core, so a lockfile grep shows something like vite@npm:@voidzero-dev/[email protected]. Check the vite-plus version instead, and move to 1.5.1 or later.
  • npm audit does not flag this. With no CVE and no entry in the advisory database, npm audit stays quiet about it — in our test it reported only an unrelated picomatch issue. Don't treat a clean audit as a clean bill here.

Check whether the dev server is exposed

This is the half that decides whether the version matters. Look for the exposure settings in your scripts and config, and confirm what the running server is actually bound to.

Shell
# --host in package scripts, Dockerfiles, and compose filesgrep -rnE --exclude-dir=node_modules \  --include=package.json --include='Dockerfile*' --include='*compose*.y*ml' \  -- '--host([ ="]|$)' .# server.host set in a Vite configgrep -rnE --exclude-dir=node_modules --include='vite.config.*' 'host *:' .# A running dev server bound to anything but loopback is reachabless -ltnp | grep -E ':(517[3-9]|6006|4321|3000)\b'

In the ss output, a Local Address of 0.0.0.0 or * means the server is listening on every interface; 127.0.0.1 means loopback only. (lsof -nP -iTCP:5173 -sTCP:LISTEN reads the same thing where ss isn't available.) One caution: server.host inside a hmr: { ... } block sets the HMR socket's host, not the server's — match the host: directly under server, not every host: in the file.

Check frameworks that run Vite's dev server

A framework that embeds Vite inherits this bug when it runs a vulnerable Vite dev server and exposes it. The clearest case we checked is Storybook: its Vite builder mounts Vite's dev middleware at the server root, so the same /@id/ and /@fs/ request handling is reachable through Storybook. Until the October 12, 2026 releases (8.6.19, 9.1.21, 10.6.2, and 11.0.0-alpha.6), storybook dev bound to 0.0.0.0 by default; those releases changed the default to 127.0.0.1. So an older Storybook, on a vulnerable Vite, on a shared network is exactly the exposed case. Check each framework's version and the host it binds, and update both the framework and the Vite under it:

Shell
# The resolved Vite under a framework, and the framework's own versionnpm ls vite storybook nuxt @sveltejs/kit astro --all 2>/dev/null

Nuxt and SvelteKit embed Vite's dev server too — Nuxt shipped its own dev-server exposure fixes this autumn, and SvelteKit a related preview-server fix. Treat any framework dev server you put on the network as carrying the Vite bug until you've confirmed its resolved Vite is patched.

How to fix

Update Vite

Move to the fixed release on your line — 8.3.7, 8.2.6, 8.1.7, 7.3.8, or 6.4.6 — or to the latest, 8.3.8. On 8.2, use 8.2.6: the 8.2.5 that the advisory names was never published to npm, so npm install [email protected] fails with ETARGET. Then restart the dev server: a running Vite process keeps serving the old code until it's restarted, and HMR doesn't reload Vite itself.

Shell
npm install vite@latest --save-dev     # or your line's fixed versionnpm ls vite rolldown --all             # confirm both moved

On Vite 8, updating Vite moves its Rolldown with it; the second command confirms Rolldown is at 1.2.18 or later (1.1.7 or later on Vite 8.1). A fresh install today gets 1.2.19 or 1.1.8. Don't pin Rolldown yourself to chase the fix. For Vite+, update the vite-plus package to 1.5.1 or later. For a framework, update the framework too, so the Vite it resolves is the patched one.

If you can't patch immediately

Take away the exposure, which is a complete mitigation here — unlike many advisories, the vendor's own condition says an un-exposed dev server isn't affected at all. Bind the server back to localhost:

  • Drop --host from your dev script and any CMD/command in Dockerfiles and compose files, and remove server.host (or set it to 'localhost') in vite.config.
  • Don't publish the dev port. If you must reach the server from another machine, use an SSH tunnel (ssh -L 5173:localhost:5173 user@host) rather than opening the port.
  • In containers, bind the published port to the host's loopback — -p 127.0.0.1:5173:5173 — so only the host can reach it.

If a vulnerable server was exposed

If a dev server ran a vulnerable Vite while reachable on an untrusted network, treat the secrets it could read as exposed. The file-read half reaches anything the dev process can read, and exposed Vite dev servers are actively scanned for exactly those files. Rotate what's in reach: the project's .env files, any cloud credentials and tokens on the box, and secrets in the developer's home directory such as ~/.aws/credentials. Code execution means the machine itself should be treated as untrusted until it's checked.

End-of-life versions

Vite ships security fixes for its current minor, the previous major's latest minor, and the lines just behind them — which is why 8.3, 8.2, 8.1, 7.3, and 6.4 got releases and nothing older did. The advisory names 4.0.0 as the oldest affected version, but the only fixed release below 7.x is 6.4.6, so every line from 6.3 down to 4.0 — and the 5.x series in between — is fixed only by upgrading onto the 6.4 line or newer. The resolver's unescaped-name pattern predates 4.0.0, so older installs should be treated as affected and moved regardless of where the advisory draws its line.

For most projects the cleanest target is the current line, 8.3 (8.3.8), which also moves you onto the maintained Rolldown-based build. A 5.x or 4.x app faces a major upgrade either way; crossing to 8.x once is less work than doing it twice. Until you can, the dev server for an unsupported Vite is only exposed if you expose it — keep it on localhost and the bug has nothing to reach.

Timeline and exploitation status

All times are US Central: CST (UTC−6).

  • Fix
  • Advisory or rating
  1. Nov 18
    1. 1:06 AMVite announces the advisory on Bluesky and asks users to update

  2. Nov 9
    1. 7:35 PMVite publishes GHSA-p4f5-76q4-ffq7, rated Critical, with a proof of concept

    2. 8:00 AMVite+ 1.5.1 publishes, bundling the fixed Vite and Rolldown

    3. 5:00–5:06 AMVite 8.2.6, 7.3.8, 8.1.7, and 6.4.6 publish to npm. 8.2.5, released on GitHub 10 seconds before 8.2.6, never reaches npm

    4. 4:25 AMVite 8.3.7 publishes to npm, pinning the fixed Rolldown

    5. 4:05–4:19 AMRolldown 1.1.7, 1.2.18, and 1.3.1 publish to npm (1.2.17 went to crates.io only)

    6. 2:42 AMRolldown merges the escaping fix to its main, 1.2, and 1.1 branches (#11584 (opens in a new tab), #11585 (opens in a new tab), #11586 (opens in a new tab))

The advisory includes a proof of concept, which we haven't tested. GHSA-p4f5-76q4-ffq7 isn't listed in CISA's Known Exploited Vulnerabilities catalog as of 11:30 AM CST, November 18, 2026.

Updates to this post

We'll revise this post and mark it updated when any of these change: a CVE ID or an NVD score for the advisory; a KEV listing; a report of exploitation or a public proof-of-concept repository; a change to Vite's advisory; or fixes from the frameworks that embed Vite's dev server.

Get help

An outside scan can't read your lockfile, but it can see what puts this bug in reach: a dev server answering on the internet. Our free security audit checks the domains and public IPs you verify for exposed services, publicly visible software versions with known CVEs, TLS, email security, DNS, and HTTP headers.

Atlas Service
Free Security Audit
Non-destructive checks of your verified domains and public IPs, and a PDF report with every finding ranked and a fix for each, within 3 business days of verification.

Sources

Atlas Security
Security Advisories

The Atlas security byline: write-ups of vulnerabilities in widely deployed software — what changed, who is exposed, and what to check — verified against primary sources.

Continue reading

More from the Atlas Meridian

Security
11 min read

OpenSSH's November 2026 Password Authentication Bypass on musl: 10.6p2

OpenSSH 10.6p2 fixes a critical password-auth bypass that lets a remote attacker log in to musl systems without PAM, such as Alpine. No CVE. Who's exposed and what to do.

Atlas Security
Security Advisories
View category
Security
13 min read

APT's November 2026 Download Verification Bypass: CVE-2026-109518

An APT flaw lets an on-path attacker swap a verified package and run code as root on Debian, Ubuntu, and derivatives. Debian and Ubuntu have fixes out.

Atlas Security
Security Advisories
View category
Security
15 min read

Rails' November 2026 Active Storage File Write and Read: GHSA-f3w6-w445-fwh9

Rails 8.1.4.1 and 8.0.5.2 fix a Critical (CVSS 9.8) Active Storage flaw: an uploaded image can make ImageMagick write, and read, files on the server.

Atlas Security
Security Advisories
View category
Engage · Atlas Engineering

Reading is one thing. Shipping is another.

If this writeup resonates with what you're building, let's talk. We'll connect you directly with an engineer to assist.