NoisyTCP CVE-2026-99288: What Linux Admins Need to Know

If your box has a TCP port open to the internet, and almost every Linux server does, NoisyTCP is the kernel bug you patch this week. Tracked as CVE-2026-99288, this 11-year-old Linux kernel flaw lets unauthenticated attackers crash Linux servers remotely, just by sending crafted TCP packets at Nginx, Apache or plain old SSH. No login, no password.

·
Published November 17, 2026 · Updated November 17, 2026
·
⏱ 7 min read

Highlights

1 NoisyTCP is a race condition in the Linux TCP networking stack, specifically the lockless established-socket lookup used by both IPv4 and IPv6. Researcher Hyunwoo Kim (@v4bel) reported and patched it.
2 The bug came in with commit 079096f103fa in October 2015 and first shipped in Linux 4.4, so it sat in the TCP stack for roughly 11 years.
3 The fix, "inet: reject unhashed sockets during established lookups" (commit 4e2b3ef19e0e), hit the networking tree on November 3, 2026 and mainline on November 5.
4 Fixed stable kernels are 6.12.117, 6.18.63 and 7.3.4, plus 7.4-rc2. If you run the 7.3 kernel series, 7.3.4 is your target.
5 On arm64, POWER and RISC-V, those releases carry only half the fix. A second memory-ordering patch (5b5d8f351dd9) is also needed.
6 Demonstrated impact is a kernel panic. No remote code execution has been shown, no public exploit code exists, and it is not in CISA KEV.

♦ BLACK FRIDAY HOSTING DEALSLooking for reliable Linux VPS and cloud hosting? Compare exclusive deals from Cloudways, Verpex and DigitalOcean and more.♦ VIEW ALL DEALS →

The fix went in quietly in early November, the CVE record followed on November 10, and the full write-up dropped on November 15 once the linux-distros embargo ended. That gap is why NoisyTCP patch status is so uneven right now. Debian, Arch and Oracle have shipped fixed kernels, Ubuntu, RHEL and Fedora haven't, and vendors still disagree on severity. Status below was checked on November 17, 2026.

CVE ID CVE-2026-99288 (NoisyTCP)
Severity Red Hat: Important, draft CVSS 7.5 · Oracle: Important, 8.1 · Canonical: Medium · NVD: not scored
Fixed Kernels 6.12.117 · 6.18.63 · 7.3.4 · 7.4-rc2 (arm64, POWER, RISC-V also need 5b5d8f351dd9)
Introduced Linux 4.4, October 2015 (commit 079096f103fa)

What NoisyTCP Actually Does to Your Kernel

Think of the kernel's list of open connections as a phone book that many CPU cores read at the same time, without locking it. Entries get erased and rewritten fast. So one core can look up a half-open connection while a second core erases it and, through the SYN cookie path, writes a brand new entry at the same memory address with the same addresses and ports.

The lookup double-checked the addresses and ports but never asked whether the entry was still actually in the phone book. The new entry has no listener attached, the kernel reads a NULL pointer inside tcp_v4_rcv(), and the whole host panics. The same unchecked lookup exists in the IPv6 path, which is why the fix touches both.

That makes it a pre-auth remote DoS. The attacker doesn't need credentials or any app-level input, just ordinary connection attempts mixed with IPv4 fragments. Fragments hash differently from normal TCP packets, which spreads the work across receive queues and gives the race two CPUs to play with.


Remote denial of service "can be achieved without much difficulty."


Hyunwoo Kim, NoisyTCP advisory, November 2026

Timing matters here. On QEMU/KVM guests the race was won in about 5 seconds to 2 minutes, because vCPUs preempt each other constantly. On a quiet bare-metal host with 8 RX queues it took around 14 hours. So if your stack lives on a Linux VPS, you are in the easy-to-hit group, not the hard one.

One more thing. If you patched last week's TCP bug, CVE-2026-98432, that does not cover you here. NoisyTCP is a separate race in the same code path. This is also the second big kernel bug from the same researcher this year, after the Dirty Frag privilege escalation in May.

NoisyTCP Patch Status: Which Kernels and Distros Are Exposed

You need all of these for the bug to bite: a kernel without commit 4e2b3ef19e0e, a TCP listener reachable from an untrusted network, and two or more receive contexts. Those are either multiple RX queues on the network card or RPS, a kernel feature that hands incoming packets to several CPUs. Most cloud instances have both on by default.

Watch the version numbers, though. Debian 12 ships a patched 6.1.192, while the upstream 6.1.192 release doesn't have the fix at all. Distro kernels carry their own backports, so trust the vendor changelog, not the number. Here's where things stood on November 17:

Distribution Kernel / Advisory Status Notes
Debian 13 "trixie" and Debian 12 "bookworm" 6.12.117-1 (DSA-6628-1) · 6.1.192-1 (DLA-4906-1) Fixed sid fixed in 7.3.4-1
Arch Linux linux 7.3.4 · linux-lts 6.18.63 Fixed arm64 users still need the second patch
Oracle Linux 7 to 10 (UEK) ELSA-2026-500630 to 500632 Fixed Shipped November 12 and 13
Rocky Linux, AlmaLinux No advisory yet In Progress CIQ backports proposed for Rocky
Ubuntu 16.04 to 26.10 No fixed kernel Needs Evaluation Canonical priority: Medium
RHEL 8, 9, 10 No erratum yet Affected RHEL 6 and 7 not affected
Fedora 45 7.3.3 in updates Affected Watch for 7.3.4 or later
Debian testing "forky" 7.2.15-1 Affected 7.2 branch gets no fix
SUSE, Amazon Linux No advisory found Unknown Watch vendor trackers

Ubuntu users are the big waiting group, and that includes fresh installs of Ubuntu 26.04 LTS. The upstream 6.6, 6.1, 5.15 and 5.10 LTS lines also have no fixed release yet. The patch is queued for them, but not out.

How to Check If Your Server Is Vulnerable

Here's the short version, in order:

  1. Find your network interface name with ip -br link.
  2. Check your kernel version and CPU architecture.
  3. Check the vendor changelog for CVE-2026-99288. This is the only check that really counts.
  4. Count your RX queues and check whether SYN cookies are on.

The examples below use eth0. On cloud VMs yours might be ens5 or enp0s3, so swap it in.

$ ip -br link
lo UNKNOWN 00:00:00:00:00:00 <LOOPBACK,UP,LOWER_UP>
eth0 UP 52:54:00:a1:b2:c3 <BROADCAST,MULTICAST,UP,LOWER_UP>
$ uname -r
$ uname -m
6.12.117+deb13-cloud-arm64
aarch64

That output is a trap. The version number says fixed, but aarch64 means this box still needs the second memory-ordering patch. On x86_64 you'd see x86_64 instead.

Now the check that actually matters: does your distro's kernel changelog mention the CVE? On RHEL, Rocky, Alma and Fedora:

$ sudo dnf -q updateinfo list --cve CVE-2026-99288
$ rpm -q --changelog kernel-core-$(uname -r) | grep -c CVE-2026-99288
0

The first command printed nothing, which means no advisory exists yet, and a count of 0 means your running kernel doesn't carry the fix. On Debian, run zgrep -c CVE-2026-99288 /usr/share/doc/linux-image-$(uname -r)/changelog.gz. On Ubuntu, sudo pro fix CVE-2026-99288 asks Canonical's data directly and tells you whether a fix exists.

Next, the race conditions. Combined above 1 means multiple RX queues, basically separate lanes your CPUs use to pull in traffic.

$ ethtool -l eth0
Channel parameters for eth0:
Pre-set maximums:
RX: n/a
TX: n/a
Other: n/a
Combined: 8
Current hardware settings:
RX: n/a
TX: n/a
Other: n/a
Combined: 4

SYN cookies are a SYN flood defence: under pressure, the kernel encodes half-open connection info into the reply instead of storing it. That cookie path is what the race abuses.

$ sysctl net.ipv4.tcp_syncookies
net.ipv4.tcp_syncookies = 1

If a box already rebooted on you out of nowhere, check the previous boot's kernel log. Fair warning: a panic in interrupt context often dies before anything reaches disk, and Debian doesn't keep a persistent journal by default. Example output when it does get logged:

$ sudo journalctl -k -b -1 | grep -iE 'NULL pointer|Fatal exception'
kernel: BUG: kernel NULL pointer dereference, address: 0000000000000012
kernel: Kernel panic - not syncing: Fatal exception in interrupt

The address 0x12 is the giveaway. For production boxes, kdump, pstore or a serial console will catch the trace far more reliably than the journal.

Temporary Mitigations and Workarounds If You Can't Reboot Yet

The real fix is a kernel with 4e2b3ef19e0e (plus 5b5d8f351dd9 on arm64), then a reboot. On Ubuntu the meta package is linux-image-generic. On Debian it's linux-image-amd64, or linux-image-cloud-amd64 on cloud images.

$ sudo apt update && sudo apt install --only-upgrade linux-image-generic && sudo reboot
Reading package lists... Done
Building dependency tree... Done
linux-image-generic is already the newest version (6.8.0-88.89).
0 upgraded, 0 newly installed, 0 to remove and 0 not upgraded.

See that "already the newest version" line? That's what most Ubuntu admins will get today, because no fixed package exists yet. On Fedora 45 and RHEL-family boxes it looks similar right now:

$ sudo dnf update kernel
Updating and loading repositories:
Repositories loaded.
Nothing to do.

So you need a workaround. Option one is turning off SYN cookies, which the tcp(7) man page documents under tcp_syncookies.

$ sudo sysctl -w net.ipv4.tcp_syncookies=0
$ echo 'net.ipv4.tcp_syncookies=0' | sudo tee /etc/sysctl.d/99-noisytcp.conf
$ sudo sysctl --system
net.ipv4.tcp_syncookies = 0
net.ipv4.tcp_syncookies=0
* Applying /etc/sysctl.d/99-noisytcp.conf ...
net.ipv4.tcp_syncookies = 0

Trade-off: you lose SYN flood protection. Option two is squashing traffic onto one receive queue and turning RPS and RFS off. That kills the race too, but throughput drops.

$ sudo ethtool -L eth0 combined 1
$ for q in /sys/class/net/eth0/queues/rx-*/rps_cpus; do echo 0 | sudo tee $q; done
$ for q in /sys/class/net/eth0/queues/rx-*/rps_flow_cnt; do echo 0 | sudo tee $q; done
$ sudo sysctl -w net.core.rps_sock_flow_entries=0
0
0
net.core.rps_sock_flow_entries = 0

After combined 1 there's only one queue left, so each loop prints a single 0. Some drivers want rx 1 tx 1 instead of combined 1. And ethtool -L doesn't survive a reboot, so persist it with a udev rule or your network manager's ethtool settings if you need it to stick.

Common mistake: setting tcp_syncookies=0 on the host and calling it done. That sysctl is per network namespace, so every container with its own netns still has cookies on and is still exposed. Set it per container, for example docker run --sysctl net.ipv4.tcp_syncookies=0 ..., or in Kubernetes through securityContext.sysctls, where it counts as a safe sysctl. Then confirm with sudo nsenter -t <PID> -n sysctl net.ipv4.tcp_syncookies.

$ docker run -d --name web --sysctl net.ipv4.tcp_syncookies=0 nginx
$ docker exec web cat /proc/sys/net/ipv4/tcp_syncookies
3f9c2a7d1e4b8c6f0a2d5e7b9c1f3a6d8e0b2c4f6a8d0e2b4c6f8a0d2e4b6c8f
0

Running a fleet? One Ansible ad-hoc call shows you everything at once: ansible all -b -m shell -a "uname -rm; sysctl -n net.ipv4.tcp_syncookies; ethtool -l eth0 | tail -1".

Neither workaround fixes the lookup itself. Once you're on a fixed kernel, roll back: sudo sysctl -w net.ipv4.tcp_syncookies=1, delete /etc/sysctl.d/99-noisytcp.conf, and restore your queue count.

What Is Still Unresolved

Four things. First, the 6.6, 6.1, 5.15 and 5.10 LTS lines have no upstream fixed release yet, which is a lot of long-lived enterprise kernels. If you're stretching an old RHEL box through extended life cycle support, watch your vendor closely.

Second, arm64. The memory-ordering patch hasn't reached the 6.12.117, 6.18.63 or 7.3.4 releases, so Graviton and Ampere servers on "fixed" kernels may still be exposed until the next stable round.

Third, the severity fight over this Linux kernel TCP vulnerability. Red Hat says Important, Oracle scores it 8.1, Canonical says Medium, and NVD hasn't scored it. The argument is about how hard the race is, not what happens when it wins. A crashed host is a crashed host.

Fourth, Kim describes a second trigger path that can lead to a use-after-free. He calls it hard to hit and says code execution isn't on the table for now. Nobody has fully ruled it out, though, and live patch availability for this CVE hasn't been announced either. If memory-safety bugs in the kernel interest you, we've covered Rust in the Linux kernel before.

Questions I Get Asked About This All the Time

Is NoisyTCP remote code execution?

No. Red Hat says it causes a kernel panic but doesn't allow data changes or code execution. It's a crash bug, but a nasty one.

My kernel is 6.12.117 on arm64. Am I fixed?

Only halfway. You also need patch 5b5d8f351dd9, which hadn't landed in that release. Check your vendor changelog for both.

I patched CVE-2026-98432 last week. Does that cover NoisyTCP?

No. Different race, different fix. Check for CVE-2026-99288 on its own.

Does Fedora 45 have the NoisyTCP fix yet?

Not as of November 16. The updates repo had 7.3.3, which lacks it. Watch for 7.3.4 or newer.

My server sits behind a load balancer. Am I safe?

A TCP-terminating proxy cuts exposure a lot, but the backend listener can still be reached if anything untrusted gets to it. Don't skip the patch because of the LB.

I installed the new kernel but uname still shows the old one. Did it work?

Not until you reboot. The running kernel is what counts.

Does this hit my desktop?

Only if it has a listening port reachable from an untrusted network. Most laptops don't, but check anyway. We covered a similar question with an earlier Ubuntu desktop flaw.

Short version: check the changelog, not the version number. Patch and reboot where you can, use the sysctl workaround where you can't, and keep an eye on Ubuntu, RHEL, Fedora and arm64 updates over the next few days.

LinuxTeck - A Complete Learning Blog

Tech News Stay updated with the latest Linux and open-source news, covering new releases, distro updates, security patches, and enterprise developments, delivered in plain language for sysadmins and developers.

About John Britto

John Britto Founder & Chief-Editor @LinuxTeck. A Computer Geek and Linux Intellectual having more than 20+ years of experience in Linux and Open Source technologies.

View all posts by John Britto →

Leave a Reply

Your email address will not be published.

L