Ten Node CVEs, one invisible runtime
Node.js fixed ten CVEs on October 21, including a Brotli decompression RCE and a dns.resolve NUL bug that defeats SSRF allowlists. A rebuild runbook for Node in Docker, why upgrading isn't enough for snapshots and SEAs, and the honest part: image scanners can't see the runtime, and on release day neither did Pinhaul.
Pinhaul Crew
October 23, 2026 · 8 min read
Node.js shipped its October security releases on Wednesday, October 21: ten CVEs across the three supported lines, fixed in v26.12.1, v24.22.1 and v22.23.4. No CVSS scores are published yet and no one is reporting exploitation in the wild — but one of the HIGHs is remote code execution in Brotli decompression, and the roads in are ordinary traffic: request bodies, fetch() responses, WebSocket frames. If you run Node anywhere in Docker, today's move is simple: rebuild on a fixed release — and rebuild anything you baked from it, because one of these CVEs survives a runtime upgrade. Below is the runbook, plus the part we have to say out loud: the runtime in official node images is invisible to image scanners, and on release day our own product didn't flag it either.
We are Pinhaul; this blog is our field notes. PHP2Shell was fixed the same day these Node releases landed, which turned this week into a live test of what we shipped after it. Part of it passed. Part of it was the same hole again. Both halves are below.
What shipped
One release per supported line, ten CVEs between them, severity labels from Node itself and no CVSS scores published as we write. Node 20 and older are end-of-life and get no fix — and so does the odd-numbered line 25. If you're on any of those, this is a migration post, not a patch post.
| Branch | Fixed in |
|---|---|
| Node 26 | 26.12.1 |
| Node 24 | 24.22.1 |
| Node 22 | 22.23.4 |
| Node 20 and older | no fix — end-of-life; move to 22.23.4+, 24.22.1+ or 26.12.1+ |
| Node 25 (odd-numbered line) | no fix — end-of-life; move to 26.12.1+ |
The one to patch for first: Brotli RCE
CVE-2026-111497 (HIGH) is an out-of-bounds write in the Brotli decoder Node bundles. Decompress attacker-controlled Brotli data and the result is remote code execution. The catch is how many places decompress without being asked:
- HTTP request bodies with
Content-Encoding: br— inbound, on any service that accepts compressed request bodies. fetch()responses from a hostile server — fetch decompresses automatically, so a URL your app calls is enough; you never chose to decode anything.- WebSocket frames compressed with Brotli.
Affects 26, 24 and 22, and there is no upstream workaround. Until you can rebuild, the interim is to keep attacker-controlled Brotli away from Node entirely: strip or reject Content-Encoding: br on untrusted inbound traffic at the reverse proxy, and disable Brotli decompression in the app where possible. That's a wall in front of one door — the door stays open behind it.
The name-checking bugs
Four CVEs in the machinery that answers "is this host the host I meant?" — the checks SSRF allowlists and TLS verification lean on:
- CVE-2026-105482 (HIGH, 26/24/22):
dns.resolvetruncates a name that contains a NUL character, so Node can resolve a different name than the one the application checked. Hostname-based SSRF allowlists stop being allowlists: you approved one name, the resolver went to another. Interim: reject any hostname containing a NUL before resolving it. - CVE-2026-106069 (HIGH, 26/24/22): TLS certificate hostname verification can be bypassed when the host loosely resolves to an IP address.
- CVE-2026-105483 (HIGH, 26/24 only): the same family of bypass when the IP address is written in fullwidth digits — the wide Unicode look-alikes of 0-9.
- CVE-2026-106072 (MEDIUM, 26/24/22): non-canonical IPv4 hostnames — octal or hex spellings of the address — let a certificate's DNS SANs authenticate an IP endpoint.
The working summary for these: on an unpatched Node, "I verified the hostname" verified less than it says. One more in the middle tier: CVE-2026-105481 (MEDIUM, 26/24/22) — a keep-alive socket can process a response that was written before the request was sent. On a reused connection that is the wrong answer attached to the right request; behind a proxy, that's request-smuggling and cache-poisoning territory.
The one where upgrading isn't enough
CVE-2026-106074 (MEDIUM, 26/24/22): the entropy caches behind crypto.randomUUID() and crypto.randomInt() get saved into user-land startup snapshots and single-executable applications (SEAs). Every process launched from the same snapshot or SEA repeats the same supposedly-random values — same UUIDs, same "random" ints, every run, everywhere it runs.
Which means upgrading the runtime is step one of two. Rebuild your snapshots and SEAs on the fixed version, and rotate any IDs, tokens or secrets those builds generated. If a SEA was minting session tokens or API keys, treat everything it minted as guessable — because from the outside, that's how it behaved.
The rest, plus what rode along
Three LOWs, two of them Node 26 only: CVE-2026-106070 returns attacker-controlled prefix bytes after failed ZipFile header validation; CVE-2026-106071 over-grants filesystem access in the Permission Model through a wildcard radix lookup; CVE-2026-106073 bypasses the Permission Model via fs.realpathSync and affects 26, 24 and 22. On unpatched versions, don't lean on --permission as a security boundary.
The releases also update bundled dependencies: c-ares, OpenSSL (24/22), SQLite (22), Undici, and CA certificates — the last one following the eMudhra intermediate CA compromise. If you wanted a reason to rebuild beyond the CVE list, a fresh trust store is a decent one.
Threat reality, briefly
No CVSS scores published yet. No in-the-wild exploitation reported as we write. No CISA alert. Genuinely quieter than PHP2Shell's day one — and quieter is not safer: the Brotli bug is a no-workaround RCE, so patch on an RCE schedule, not on the calendar the current noise suggests.
The patch runbook, Docker edition
There is exactly one fix: run a fixed Node. Everything below is the order of operations.
- Find every Node. In Docker, the official images declare what they are:
Then rundocker ps -q | xargs docker inspect --format '{{ .Name }} {{ range .Config.Env }}{{ println . }}{{ end }}' | grep NODE_VERSIONnode -vinside the containers that actually serve. Then look everywhere Docker doesn't: CI images, base images you publish for other teams, Lambda and other serverless runtimes. Electron ships its own Node and is not covered by this release — track it separately. - Bump the base image tag and pin by digest. Builds on
node:22go tonode:22.23.4, 24 to24.22.1, 26 to26.12.1— and pin by digest so the tag can't move under you later. If you publish base images for other teams, rebuild the whole chain. - Rebuild — don't just restart. The runtime is baked into the image. A restarted container is the same old binary with fresh uptime; rebuild the image on the fixed tag, then recreate the containers from it.
- Rebuild snapshots and SEAs on the fixed version. Per CVE-2026-106074, the flaw lives in the artifact, not only in the runtime it was built with.
- Rotate what affected SEAs and snapshots generated — IDs, tokens, secrets. They were not random. Rotate them.
- Interim, if rebuilding takes days: strip or reject
Content-Encoding: bron untrusted inbound traffic at the proxy, disable Brotli decompression in the app where you can, and reject NUL-containing hostnames before resolving. - Verify.
docker run --rm <image> node -vin CI, or open the container in Pinhaul. Trust the version string inside the serving container over any scanner — ours included.
What the scanner sees: nothing
The blind spot from the PHP2Shell post is not a PHP thing; it's an official-language-image thing. The Node in node:* images is compiled from source, so it exists in no OS package database, and image scanners — Trivy included, which is what our agent runs — have no package record to match an advisory against. A vulnerable node:22.23.3 image scans green where the runtime is concerned.
And in the interest of exactness: on release day, Pinhaul didn't flag this either. The runtime advisory table we shipped after PHP2Shell had exactly the PHP2Shell entries in it. Ten Node CVEs existed, our product said nothing, and nothing in the product was going to notice until a person updated the table by hand.
Two runtimes in one week, same lesson: a green image scan says nothing about the runtime baked inside it.
What we shipped today
- All ten CVEs are in the runtime advisory table, matched per release line: a container on 22.23.3 gets the 22.23.4 fix, the 24 line points at 24.22.1, the 26 line at 26.12.1, and EOL lines (20, 25) are told to move to a supported line rather than pointed at a patch that will never come.
- The general fix, so nobody hand-types the next one: Pinhaul now syncs nodejs.org's official release index every few hours, and any container whose Node is behind its line's latest security release gets a finding the same day — before anyone writes the advisory up, before any scanner database carries the CVEs. Existing containers are re-checked the moment new data lands, without waiting for the next agent scan. Customers need to do nothing: the agent already reports NODE_VERSION.
What this still doesn't see
The limits, plainly: this reads the version official images declare. Node from distro packages — apk add nodejs, apt install nodejs — declares nothing, but Trivy covers it once the distro feeds update. Node installed some other way — nvm, a copied binary — with no NODE_VERSION is not seen yet. And the Python SDK doesn't see Node at all. The long tail stays long.
Where we're reading
- The official Node.js advisory — the source of record for the CVE list, affected versions and upgrade paths.
Rebuild, don't restart. Rebuild the snapshots and SEAs too, rotate what they minted, and trust the version string inside the serving container over any scanner — ours included. If that's the kind of runtime-aware watching you want on your stack, the waitlist is here.
Want this watching your stack?
Every engine in this story runs nightly for the teams on Pinhaul. Join the waitlist — we onboard a few builders every week.