DEV Community

sunah kim
sunah kim

Posted on Originally published at zenn.dev

30 Minutes After a CVSS 9.8 Alert — The Conclusion Was 'No Patch Yet' and a 50-Second Reboot

Intro

A post went around on social media: "Linux kernel TCP (IPv4) use-after-free, CVE-2026-98432, CVSS 9.8. If your cloud server has open TCP ports, update immediately." I run a small production server (Ubuntu 24.04, one VM, Caddy reverse proxy + Java app on Docker), so here's what I actually checked, in order, and where I got stuck.

Spoiler: no fixed kernel existed yet for my kernel flavor. Instead, I applied other security patches that had been sitting there waiting for a reboot (about 50 seconds of downtime), and left a one-liner that will tell me the moment a fix ships.

The setup

flowchart LR
    U["Browser"] -->|"443"| C["Caddy container<br/>reverse proxy"]
    C -->|"8080"| A["app container<br/>Java 17"]
    A --> D["Managed DB<br/>(outside the VM)"]
    subgraph VM["1 VM · Ubuntu 24.04 · open ports 22/80/443"]
      C
      A
    end

The important part: no data lives on this server. The DB and file storage are external, so the box is disposable — which made the later reboot decision easy. If the DB had been on the VM, I'd have started with a snapshot.

Pitfall 1: The server I meant to inspect wasn't there

SSH to the IP in my notes: Operation timed out. Not just 22 — 80 and 443 were closed too. Yet the service domain happily returned 200. The service had been migrated to a different VM (new IP, different SSH key), and my notes and scripts still had the old one.

dig +short app.example.com   # where the domain points NOW
for p in 22 80 443; do nc -z -w 5 203.0.113.10 $p && echo "$p open" || echo "$p closed"; done
Enter fullscreen mode Exit fullscreen mode

Step zero of any security check is "which box is the service actually running on?" Don't trust an IP somebody (or an old doc) gave you — walk backwards from the domain. Then fix the notes the same day; the next person losing ten minutes to a timeout is you.

Pitfall 2: "CVSS 9.8" and "a patch exists for my server" are different sentences

The post said "patches are rolling out for Ubuntu 22.04/24.04, run apt upgrade." But the official Ubuntu CVE tracker showed, for my release:

Kernel package Status
linux (6.8 GA) Vulnerable, work in progress in 6.8.0-149
linux-aws (6.8) Vulnerable, work in progress
linux-aws-7.0 (this server) Vulnerable — no target version even listed

So running apt upgrade today would not fix this CVE at all. The post wasn't wrong, exactly — "rolling out" just doesn't mean "has reached every kernel flavor." Cloud kernels (-aws/-azure/-gcp) and HWE series often lag behind the GA kernel.

Finding out which flavor you're on:

uname -r                                  # e.g. 7.0.0-1012-aws → "aws" flavor, 7.0 HWE
dpkg -l 'linux-image-*' | awk '/^ii/{print $2, $3}'
apt changelog linux-image-7.0.0-1014-aws 2>&1 | grep -i -m3 'CVE-2026-98432' || echo "(not mentioned)"
Enter fullscreen mode Exit fullscreen mode

One detail: don't silence the changelog command with 2>/dev/null. If you do, you can't tell "empty because the fetch failed" from "empty because the CVE isn't mentioned."

Pitfall 3: "Everything is updated" — but the running kernel is old

apt list --upgradable showed 0 packages. Yet uname -r said 1012, kernel 1014 was installed, and /var/run/reboot-required was still waving its "System restart required" flag. Someone had installed the patches and then not rebooted for a week.

Replacing kernel files on disk changes nothing about the kernel running in memory. Libraries take effect on process restart; the kernel alone requires a reboot to be applied.

cat /var/run/reboot-required 2>/dev/null || echo "no reboot needed"
cat /var/run/reboot-required.pkgs 2>/dev/null   # which packages demand it
Enter fullscreen mode Exit fullscreen mode

The 50-second reboot — reasoning and timeline

Three reasons the reboot was safe: (1) both containers use unless-stopped, so they come back with Docker; (2) no data on the box; (3) nothing downstream cascades if this server stops.

timeline
    title Inside the 50 seconds (measured)
    10:43:17 : sudo reboot sent
    10:43:41 : SSH back · uname -r shows new kernel
             : both containers "Up 4 seconds"
             : site still 502
    10:44:05 : app "Started" log line
             : site 200 · API 401 = healthy

Right after recovery, docker ps showed both containers Up — and the site returned 502. The reverse proxy was awake, but the JVM behind it hadn't opened its port yet. Decide your "healthy means exactly these values" check in advance: for this app it's "site returns 200 AND an auth-required API returns 401" (meaning the request made it all the way to auth). Container Up ≠ service OK.

s=$(curl -s -o /dev/null -w '%{http_code}' https://app.example.com/)
a=$(curl -s -o /dev/null -w '%{http_code}' https://app.example.com/api/auth/me)
[ "$s" = 200 ] && [ "$a" = 401 ] && echo "up: site=$s api=$a"
Enter fullscreen mode Exit fullscreen mode

What to leave behind when there is no patch: a daily one-liner

Ports 80/443 can't be closed, so the only complete defense against this CVE is the patch — and I wrote that fact itself into the report. Then I left one line that catches the moment a fix ships:

ssh server 'sudo apt update -qq && apt list --upgradable 2>/dev/null | grep linux-'
Enter fullscreen mode Exit fullscreen mode

The other signal is my row on the tracker flipping to Released. The day either one changes, I re-run Pitfall 3 and the reboot procedure. That's it.

Takeaways

  • Step zero is "which server?" — walk backwards from the domain
  • A CVSS score tells you urgency; your kernel row on the tracker tells you what you can actually do
  • Installed ≠ applied — for kernels, the reboot is the application, and reboot-required is the signal
  • Container Up ≠ service OK — define "healthy means exactly these values" ahead of time
  • "No patch exists" is a valid conclusion; a daily one-liner catches the moment that changes

The biggest thing I accomplished in 30 minutes after seeing that CVSS 9.8 post was confirming that no patch existed yet. That confirmation was the report. The 50-second reboot was a bonus.

Top comments (0)