News / Security / NoisyTCP Linux vulnerability: check your kernel
Security · Vulnerability

NoisyTCP Linux vulnerability: check your kernel

NoisyTCP is an 11-year-old Linux kernel vulnerability that lets anyone crash a server through an open TCP port. Here's how to check yours in a minute.

opnF.shNov 16, 20268 min readlinuxlinux kernelcve-2026-99288
On this page 3 sections ▾

WHAT TO DO

  • Fixed kernels are out upstream and on Debian, but Ubuntu and RHEL had none as of November 15, so many admins are waiting on their distro or running a stopgap.
  • The proven impact is a kernel panic that takes the whole host down, containers included. Nobody has shown code execution, though the researcher who found it is still trying.
New to this?Ping explains the hardest part in plain words.
CVE-2026-99288
"NoisyTCP" · Linux kernel (TCP networking)
HIGH7.5
Affected
Kernels from 4.4 (2015) on, until your branch or distro ships the fix
Fixed in
6.12.117, 6.18.63, 7.3.4
Disclosed
Nov 15, 2026
Exploited
No known exploitation
Workaround
Disable SYN cookies, or cut to one RX queue with RPS and RFS off

If your Linux server boots a kernel from before November 10, it almost certainly carries NoisyTCP, the Linux kernel vulnerability that lets anyone who can reach an open TCP port crash the whole machine without logging in. Whether someone could actually pull that off against your server comes down to how many CPUs handle its incoming packets, which two commands will show you.

Quick facts

On QEMU/KVM virtual machines, researcher Hyunwoo Kim crashed default Nginx, Apache and SSH servers in 5 seconds to 2 minutes. A quiet bare-metal server with 8 RX queues held out for about 14 hours.

Patched in October? That batch fixed a sister bug, not this one, so a kernel like 6.12.113 is still exposed.

The fix reached stable kernels on November 10, five days before Kim's full write-up went public.

#How NoisyTCP crashes a server

Every TCP packet that reaches a Linux server goes through a lookup that matches it to its connection. To keep that fast, the lookup takes no locks, and the memory that holds a connection record can be handed to a new connection the moment the old one is freed. NoisyTCP lives in the gap between those two facts, and it needs two CPUs working at the same instant:

  1. CPU A starts looking up a packet's connection and keeps hold of where that connection lives in memory.
  2. CPU B, handling another packet, frees that connection, and the kernel reuses the same memory straight away for a new half-open connection rebuilt from a SYN cookie, carrying the same addresses and ports.
  3. CPU A grabs the recycled memory and sees matching addresses, but this connection was never added to the lookup table and has no listener behind it, so the kernel follows an empty (NULL) pointer and panics.
ELI5: explain it simply
Ping explainsNo experience needed.
THE WORDS
Half-open connection
A connection that has started its handshake but hasn't finished it yet.
SYN cookie
A trick that rebuilds a half-open connection from the packet itself, so a flooded server doesn't have to store it.
Kernel panic
The kernel stops the whole machine because it hit an error it can't recover from.
THE IDEA

Picture a coat check with two attendants. One reads ticket 42 and walks toward hook 42. Meanwhile the other returns that coat and hangs a brand-new coat with no name tag on the same hook. The first attendant grabs it, finds no owner, and the whole counter shuts down.

The hard part for an attacker is getting two CPUs onto the same connection at once. Kim's proof of concept does it by mixing IPv4 fragments with ordinary TCP packets, because the network card sorts the two into different RX queues. Virtual machines fall faster for a separate reason: when the hypervisor pauses a virtual CPU in the middle of a lookup, the window CPU B needs gets much wider.

For now, that crash is the whole impact: Red Hat says the flaw causes a kernel panic but doesn't let an attacker change data or run code. Kim's write-up describes a second route, though, that could cause a use-after-free (the kernel using memory after it's been handed to something else, the kind of bug that sometimes becomes a full takeover). He calls that route generally hard to trigger and says his work on a general-purpose TCP exploit is ongoing.

#Is your Linux server exposed to NoisyTCP?

Three things have to line up: a kernel without the fix, more than one CPU handling incoming packets, and a TCP port the attacker can reach. Start with the kernel. Here's the check on our own test machine, a small VM:

check your running kernel
student@opnf:~$ uname -r
6.18.44-fc-v470
# the 6.18 branch is fixed from 6.18.63, so this kernel has the bug

uname -r shows the kernel that's running now, and that's what counts: an updated kernel sitting on disk does nothing until you reboot into it. Upstream, the fix landed in three branches, and the older long-term branches still had none:

SystemStatus
6.12 before 6.12.117 (includes October's 6.12.113)Affected
6.12.117 and laterFixed
6.18 before 6.18.63Affected
6.18.63 and laterFixed
7.3 before 7.3.4Affected
7.3.4 and laterFixed
7.2 (branch ended at 7.2.15)Affected, no upstream fix
5.10, 5.15, 6.1 and 6.6Affected, no upstream fix yet

Most servers don't run upstream kernels, though. Distros copy fixes back into older branches without changing the branch number, which is how Debian 12 got a fixed 6.1 kernel when upstream 6.1 has none. So your distro's tracker is the real answer. Here's where the big ones stood on November 15:

SystemStatus
Debian 12 (bookworm)Fixed in 6.1.192-1
Debian 13 (trixie)Fixed in 6.12.117-1
Debian testing (forky, 7.2)Not fixed
Ubuntu, all supported releasesNo fixed kernel yet
Oracle Linux UEKFixed November 12 (ELSA-2026-500630)

Red Hat rates the flaw Important, with a preliminary score of 7.5 (the score on the card above), and NVD hasn't scored it yet. Cyber Kendra found Red Hat's page went up November 11, but no fixed RHEL kernel had shipped by November 15, and SUSE and Amazon Linux hadn't published advisories at all. Check Debian's tracker or your own distro's security pages before deciding you're covered, because these statuses are moving daily.

i
For your information

On Arm, POWER and RISC-V servers, ComputingForGeeks reports a second commit (5b5d8f351dd9) is also needed, and it wasn't in the fixed kernels it checked. If you run Arm in the cloud, confirm with your distro that both are in.

Can the race run on your machine?

Next, count the CPUs that can handle incoming packets. Swap eth0 for the interface that takes outside traffic (ip -br link lists them; yours might be ens4 or enp0s3):

count RX queues and RPS
student@opnf:~$ ls -d /sys/class/net/eth0/queues/rx-*
/sys/class/net/eth0/queues/rx-0
# one folder means one RX queue
student@opnf:~$ cat /sys/class/net/eth0/queues/rx-*/rps_cpus
0
# 0 means RPS (software spreading across CPUs) is off

Our test VM has one RX queue and RPS off, which is the only setup that falls outside the trigger: the kernel has the bug, but there's no second CPU to race against. If you see rx-0 through rx-3, or any rps_cpus value with a digit other than 0 (on machines with lots of CPUs, off looks like 00000000), the race can run. ComputingForGeeks found four queues were enough on its lab VMs, even with RPS off. The third condition is simpler: if the server listens on any port the internet can reach, like a website or SSH, that counts.

Containers don't change the answer. The panic happens in the host's kernel, so one exposed listener inside a container takes down the host and every other container on it.

If you're exposed

The fix is a kernel update and a reboot once your distro ships it. Until then, Red Hat lists two stopgaps, and each one costs you something:

  1. Install the fixed kernel and rebootUpdate through your package manager (Debian 12 and 13 already have it), reboot, then run uname -r again to confirm the new version is the one running.
  2. No fixed kernel yet? Turn off SYN cookiessudo sysctl -w net.ipv4.tcp_syncookies=0 closes the path that rebuilds the half-open connection. It also weakens your protection against SYN floods, and it's set per network namespace, so containers with their own namespace need it set too. And it only lasts until the next reboot: to keep it, put net.ipv4.tcp_syncookies = 0 in a file under /etc/sysctl.d/ and run sudo sysctl --system.
  3. Or take the second CPU awayCut that interface to one RX queue (sudo ethtool -L eth0 combined 1) and turn off RPS and RFS, so every incoming packet is handled in one place and nothing can race. On a busy host, that means one core does all the packet work.
!
Warning

Changing a network card's queue count with ethtool -L can drop the link on some drivers, and it doesn't survive a reboot unless your network config applies it again. Do it from a console, not over an SSH session that runs through the same interface.

There's nothing to hunt for on servers that stayed up. No attacks have been reported, and Kim hasn't released his proof-of-concept code. The one trace NoisyTCP leaves is the crash itself, so if a server on an affected kernel has recently rebooted without explanation, check its console or crash dump for a kernel panic, since it may not show up in the normal system log. The fix commit puts the crash in tcp_v4_rcv(), so that function in the panic trace is a strong hint.

If you can't tell whether a host is covered, treat any Linux server that's reachable from the internet and has more than one RX queue as exposed and pick a stopgap for it, weighing a possible crash against weaker SYN flood protection.

#Why October's patch didn't cover it

Last month's headline bug, CVE-2026-98432, is the bug Anthropic's researchers turned into remote root on servers with Multipath TCP listeners, and October's stable kernels fixed it. Both bugs trace back to the same 2015 change, which let the kernel handle half-open connections without holding the listener's lock. That fix changed when certain memory gets freed, while NoisyTCP sits earlier, in the lookup itself, so a server patched for the scarier bug is still open to this one.

On X, Kim put it more bluntly:

The era of KVM escapes is fading, and the era of TCP remote exploits has begun.
Hyunwoo Kim, on X · Source ›

Everything in this post's checks comes down to a few commands: uname -r, sysctl, and reading files under /sys. Keep our Linux commands cheat sheet handy as a quick reference for everyday commands, and the Security Pro ProCourse is where you practice server administration hands-on, in a simulated terminal in your browser. Level 1 is free.

Is NoisyTCP being exploited?
Not that anyone has reported. The catch is that Kim's technical write-up is public, so a skilled attacker could work from it without his code, which makes this quiet stretch the time to patch.
Does a firewall protect against NoisyTCP?
Only on the ports it closes. The attack needs to reach a listening TCP port, so anything your firewall lets through, like 443 on a web server or 22 if SSH is open to the internet, is still a way in.
Is my home Linux PC or Raspberry Pi at risk?
Only if someone on the internet can reach one of its TCP ports. A home machine behind a router with no ports forwarded isn't reachable that way, but one with SSH or a web server forwarded to the internet is, and it should get the update like any server.

What this means for you

Treat NoisyTCP as an outage risk, not a breach: patch and reboot when your distro ships the kernel, put a stopgap on internet-facing hosts with more than one RX queue until then, and keep an eye on Kim's research in case the harder use-after-free route turns into something worse.

opnF.sh
News that matters to the tools we teach.