
Checking 30+ Servers for the New nginx HTTP/2 RCEs in 27 Minutes with Claude Code
- CategorySecurity
- Date
- Reading8 min
On November 18, 2026, F5 shipped a security release for nginx, and the same day Wiz Research published a detailed write-up of two of the bugs in it: unauthenticated remote code execution through nginx's HTTP/2 module. nginx sits in front of almost everything I deploy, so the next evening I sat down with one question: which of my servers are actually at risk, and how fast can I fix them?
I did the whole check with Claude Code in my terminal, using the servers already listed in my SSH config. From typing the two CVE numbers to the last patch took about 27 minutes, for more than 30 hosts. Here is how it went, what I learned, and where a human still had to stay in the loop.
The two vulnerabilities in short
| CVE-2026-123633 | CVE-2026-123007 | |
|---|---|---|
| Bug | Use-after-free on an HTTP/2 connection | Integer underflow in HTTP/2 flow control, leading to an out-of-bounds copy |
| When it is reachable | HTTP/2 enabled, nginx proxying to a backend | HTTP/2 enabled, even in a default config serving static files |
| Severity | Critical, CVSS 4.0: 9.2 | Critical, CVSS 4.0: 9.2 |
| Affected open-source nginx | 1.19.7 and later | 1.9.5 and later |
| Fixed in | 1.30.7 (stable), 1.31.9 (mainline) | 1.30.7 (stable), 1.31.9 (mainline) |
The key word is HTTP/2: both bugs live in that module. If HTTP/2 is off, neither can be reached. If you cannot upgrade right away, the published workaround is to turn it off (http2 off;, and remove http2 from your listen lines), then turn it back on after upgrading. At the time of writing, no exploitation in the wild had been reported, but Wiz built working exploits in the lab, so this is a "patch now" bug.
Vulnerable is not the same as exposed
That gave me two separate questions for every server:
- Vulnerable? Is the installed nginx build missing the fix?
- Exposed? Is it vulnerable and serving HTTP/2, so the bug can actually be reached?
An exposed server gets patched first, tonight. A vulnerable but unexposed one still gets patched, just without the panic. I asked Claude to keep these in two separate columns in every report, and it made the triage obvious at a glance.
Step 1: scan everything in the SSH config
My ~/.ssh/config is the real inventory of what I manage: client servers, my demo containers at home, small VPSs, a few machines behind jump hosts, one reached through a VPN container. I asked Claude to check all of them. It read the config and ran one small, read-only script on each host over SSH, in parallel, that reported:
- the nginx binary version and the installed package version;
- whether nginx is running, and whether anything listens on port 443;
- how many
http2lines the configuration has; - any nginx (or OpenResty) running inside Docker containers, with their version, HTTP/2 lines and published ports;
- other web servers (Caddy, Apache) so a host without nginx is not mistaken for "unknown".
Nothing was changed on any server during the scan. Where my user had passwordless sudo, it also read the live configuration with nginx -T, which is more reliable than grepping files.
Step 2: don't trust the version number
The first results looked bad: every nginx was older than 1.30.7. But most of my servers run Ubuntu's nginx package, and Ubuntu does not jump to a new upstream version for a security fix. It backports the patch into the same version with a new package revision. nginx -v keeps printing the old number either way.
So the second pass checked what actually matters: the package changelog. If /usr/share/doc/nginx/changelog.Debian.gz mentions both CVE IDs, the fix is in. Claude also read the changelog of the candidate package in apt, and found that Ubuntu had already published fixed builds:
| Ubuntu release | Fixed nginx package |
|---|---|
| 26.04 | 1.28.3-2ubuntu1.13 |
| 24.04 | 1.24.0-2ubuntu7.20 |
| 22.04 | 1.18.0-6ubuntu14.23 |
That turned "upgrade nginx to 1.30.7" (a big jump, with config risk) into "install the security update from your normal repository" (routine).
The first report
About five minutes after I asked, I had a table. Host names are replaced here:
| Host | Vulnerable build? | HTTP/2 on 443 | Exposed? |
|---|---|---|---|
| srv-1 | Yes | Yes (live config, also proxies) | Yes |
| srv-2 | Yes | Yes (config files) | Probably |
| srv-3 | Yes | Yes (config files) | Probably |
| srv-4 | Yes | No | No |
| srv-5 | Yes | No | No |
| srv-6 | Yes | No (live config) | No |
| srv-7 | Yes (third-party repo, no fix yet) | No | No |
| srv-8 | Yes (host nginx stopped, nginx in containers) | No | No |
| 5 hosts | No nginx at all (Caddy or nothing) | — | No |
| 1 host | Shared hosting, patched by the provider | — | n/a |
"Probably" is an honest word here: on srv-2 and srv-3 my user had no sudo, so Claude could see http2 in the config files but not confirm the live config. It said so instead of guessing.
The scan also cleaned up my inventory. About fifteen hosts timed out or refused the connection: old client servers, machines that no longer exist, test boxes. I gave Claude the list, and it removed them from my SSH config, after making a backup. Two hosts showed a changed host key. Claude flagged them and did not touch known_hosts: a changed key is usually a reinstall, but that is exactly the warning you should never auto-accept.
Step 3: patch the exposed ones first
The three exposed servers were upgraded first with the normal security update, and Claude verified each one afterwards instead of taking it on trust:
- the installed package is the fixed build, and its changelog lists both CVEs;
- the upgrade time in
/var/log/dpkg.log, and the time nginx restarted after it; - the running nginx process uses the new binary. A package upgrade that never restarts the service leaves the old, vulnerable code in memory.
With the fix in, HTTP/2 could stay on. One detail worth knowing: on the first re-check, a plain grep -r found no http2 in /etc/nginx on srv-1 and srv-2, although the sites clearly use it. The files in sites-enabled are symlinks, and grep -r does not follow them. Claude caught the mismatch with the earlier scan and explained it. Use grep -R, or better, nginx -T.
Step 4: then the vulnerable but not exposed ones
Next came the servers that had the vulnerable build but no HTTP/2. No emergency, but "not reachable today" can change with one config edit, such as a Certbot run that adds http2 to a listen line.
- srv-4, one of my demo containers: my normal user has no sudo there, so Claude connected as root through the jump host, installed the update, ran
nginx -t, restarted nginx and checked the new version. - srv-5: patched the same way, then a full upgrade of every pending package (65 of them). Claude pointed out that the remaining systemd and Mesa updates were held back on purpose by Ubuntu's phased rollout, not errors, and that a reboot was pending. It asked before rebooting, because the machine is an LXC container and a reboot means a short outage.
- srv-8: a second look found three more nginx containers the first scan had missed, four in total, all on an older nginx image. None used HTTP/2, and three listened only on localhost, so they were not exposed, but their images need a rebuild on a patched base.
- srv-6 and srv-7 stay on the list: one is waiting for the update to show up in its apt cache, the other uses a third-party repository that had not shipped a fix yet. Neither has HTTP/2 on.
Where the human stayed in the loop
This was not "let the AI loose on production". A few moments made the difference:
- It asked before every change. Scanning was read-only. Every upgrade, restart, config edit and reboot was proposed first and run only after I said yes.
- It noticed changes it had not made. When it re-checked the exposed servers, they were already patched. Instead of claiming the credit, it reported the exact upgrade times from the logs and said plainly that it had not done them.
- It was wrong once, and corrected itself. It first told me that a duplicated host block in my SSH config was ignored. I said the second block routes that server through my VPN container. It checked with
ssh -G, which prints the settings SSH really uses, agreed I was right, and explained why: SSH takes each option from the first block that sets it, so the second block'sProxyCommandstill applies. - I gave it the context it could not know. Which hostnames are the same machine, which server is shared hosting, which old hosts can go, where I have root.
What I am keeping from this
- Your SSH config is an inventory. If you can SSH into it, you can scan it. Mine also had fifteen dead entries, now gone.
- Check the changelog, not
nginx -v. Distribution packages backport fixes without changing the upstream version. - Separate "vulnerable" from "exposed". It tells you what to fix tonight and what can wait for tomorrow morning.
- Look inside containers. An nginx in a container does not show up in
dpkgorsystemctl. - Verify the running process, not just the package. Restart after the upgrade, and check it.
- AI is a fast pair of hands, not an autopilot. The speed came from Claude doing the repetitive work across thirty machines in parallel. The safety came from it asking first, saying what it could not confirm, and me knowing my own infrastructure.
Twenty-seven minutes from a CVE number to every exposed server confirmed patched is a workflow I will reuse for the next big vulnerability. If you manage servers for your business and would like the same kind of check, get in touch.
Tags :
- nginx
- Security
- Claude Code
- AI
- DevOps
- Linux









