TL;DR — last updated November 18, 2026, 15:37 UTC
CVE-2026-99288 (CVSS 7.5, Important per Red Hat), known as NoisyTCP, lets a remote attacker with no account on the server crash a vulnerable Linux kernel through any open TCP port. No exploit code is public, and no exploitation has been reported.
This advisory is also tracked on the CloudLinux status page; subscribe there for status updates.
CVE-2026-99288 is a flaw in the Linux kernel’s TCP connection lookup. An attacker with no account on the server can crash the kernel with nothing but network traffic. Every site on the server goes down with it. The flaw affects every CloudLinux version except CloudLinux 7.
Fixed kernels for CloudLinux 7h, 8, 8 LTS and 9 LTS are in the CloudLinux beta channel, and for CloudLinux 9 and 10 in AlmaLinux’s testing repository. CloudLinux for Ubuntu 22.04 has no fix yet, and no livepatch is out. Until a server runs a fixed kernel, either of two workarounds closes the attack path: turning SYN cookies off, or limiting the network card to a single receive queue.
| Version | Affected | Status |
|---|---|---|
| CloudLinux 7 (CL7) Kernel 3.10 | No | The vulnerable code entered Linux in version 4.4. |
| CloudLinux 7h (CL7h) Kernel 4.18 | Yes | Kernel patch: 4.18.0-553.186.2.lve, beta channel Livepatch: not released yet |
| CloudLinux 8 (CL8) Kernel 4.18 | Yes | Kernel patch: 4.18.0-553.186.2.lve, beta channel Livepatch: not released yet |
| CloudLinux 8 LTS Kernel 5.14 (TuxCare ELS) | Yes | Kernel patch: kernel-lts els22, beta channel Livepatch: not available; install the fixed kernel |
| CloudLinux 9 (CL9) Kernel 5.14 | Yes | Kernel patch: 5.14.0-741.22.2.el9_9, AlmaLinux testing Livepatch: not released yet |
| CloudLinux 9 LTS Kernel 5.14 (TuxCare ELS) | Yes | Kernel patch: kernel-lts els22, beta channel Livepatch: not released yet |
| CloudLinux 10 (CL10) Kernel 6.12 | Yes | Kernel patch: 6.12.0-264.16.2.el10_3, AlmaLinux testing Livepatch: not released yet |
| CloudLinux for Ubuntu 22.04 Kernel 5.15 | Yes | Kernel patch: not released yet by Canonical † Livepatch: not released yet |
† CloudLinux for Ubuntu 22.04 takes its kernel from Ubuntu (Canonical) via apt, not from a CloudLinux stream. Ubuntu’s tracker lists the 22.04 kernel as vulnerable, with no fixed kernel released yet. Once Canonical releases one, apt update && apt upgrade and a reboot install it.
A server running an affected kernel is exposed when all three of these hold:
To check the second and third, run these as root, replacing IFACE with the name of the internet-facing interface:
sysctl net.ipv4.tcp_syncookies
ethtool -l IFACE
cat /sys/class/net/IFACE/queues/rx-*/rps_cpus
SYN cookies are on if the first command prints 1 or 2. The interface has several receive queues if ethtool shows a current Combined or RX value above 1. Packet steering is on if any rps_cpus value isn’t zero.
Until a server runs a fixed kernel, either workaround below closes the attack path. Both come from the researcher and Red Hat. KernelCare’s analysis confirmed the second one on live test servers: with a single receive queue, the attack can’t get the two CPUs it needs. As with any change to network settings, apply either one on a single server before rolling it out.
This closes the path the demonstrated crash takes. While SYN cookies are off, the server has no protection against SYN flood attacks.
sysctl -w net.ipv4.tcp_syncookies=0
echo "net.ipv4.tcp_syncookies = 0" > /etc/sysctl.d/99-cve-2026-99288.conf
The first command changes the running kernel; the second keeps the setting after a reboot. Without the file, a server that reboots comes back exposed without any warning.
The setting belongs to a network namespace. Every container with its own network stack needs it set separately, and containers started later come up without it.
To revert once the fix is installed:
rm -f /etc/sysctl.d/99-cve-2026-99288.conf
sysctl -w net.ipv4.tcp_syncookies=1
This removes the second CPU the attack needs. Every packet on the interface is then handled by one CPU, which lowers network throughput on a busy server. Changing the queue count can briefly drop the link on some physical network cards, so keep console access handy.
Before you change anything, note the current values from the check commands above, so you can set them back. Then run, replacing IFACE:
ethtool -L IFACE combined 1
for q in /sys/class/net/IFACE/queues/rx-*/rps_cpus; do echo 0 > "$q"; done
sysctl -w net.core.rps_sock_flow_entries=0
On a network card that shows separate RX channels instead of Combined, use ethtool -L IFACE rx 1. These settings don’t survive a reboot, so apply them again after each one until the fix is installed.
Unlike option 1, a container can’t undo this, because a network interface belongs to the host.
The fix will reach CloudLinux servers through the four streams below, and CloudLinux for Ubuntu 22.04 will get it from Canonical. We’ll update each stream here when its state changes.
Update: November 18, 2026, 15:37 UTC.
The same fixed kernel is now in the beta channel for CloudLinux 7h: 4.18.0-553.186.2.lve.el7h. To install it now, run:
yum update kernel --enablerepo=cloudlinux-updates-testing
reboot
After the reboot, uname -r should report 4.18.0-553.186.2.lve.el7h or a later version.
Status: November 18, 2026, 14:53 UTC.
CloudLinux has released a fixed kernel for CloudLinux 8, 4.18.0-553.186.2.lve.el8, to the beta channel. To install it now, run:
yum update kernel --enablerepo=cloudlinux-updates-testing
reboot
After the reboot, uname -r should report 4.18.0-553.186.2.lve.el8 or a later version. A fixed kernel for CloudLinux 7h hasn’t been released yet; we’ll add it here, and note when the CloudLinux 8 kernel reaches the stable channel.
Update: November 18, 2026, 15:37 UTC.
AlmaLinux has published fixed kernels to its testing repository: kernel-5.14.0-741.22.2.el9_9 for CloudLinux 9 and kernel-6.12.0-264.16.2.el10_3 for CloudLinux 10. To install one now, enable the AlmaLinux testing repository by installing its release package from AlmaLinux directly, update the kernel, then disable the testing repository and reboot.
For CloudLinux 9:
rpm -Uvh --nodeps https://repo.almalinux.org/almalinux/9/extras/x86_64/os/Packages/almalinux-release-testing-9-2.el9.noarch.rpm
dnf update 'kernel*'
dnf config-manager --disable almalinux-testing
reboot
For CloudLinux 10:
dnf install -y https://repo.almalinux.org/almalinux/10/extras/x86_64/os/Packages/almalinux-release-testing-10-2.el10.x86_64.rpm
dnf update 'kernel*'
dnf config-manager --disable almalinux-testing
reboot
After the reboot, uname -r should report a version starting with 5.14.0-741.22.2 on CloudLinux 9 or 6.12.0-264.16.2 on CloudLinux 10. AlmaLinux will move these kernels to its production repositories after community testing. Once a fixed kernel reaches the CloudLinux repositories, a regular kernel update installs it, and we’ll note that here.
Status: November 18, 2026, 14:53 UTC.
CloudLinux 9 and 10 run the AlmaLinux kernel unchanged. No fixed kernel has been released yet; Red Hat lists its RHEL 9 and 10 kernels as affected. Once a fixed kernel reaches the CloudLinux repositories, a regular kernel update and a reboot install it. We’ll add its version here.
Status: November 18, 2026, 14:53 UTC.
Fixed kernel-lts builds are in the beta channel: kernel-lts-5.14.0-284.1101.el8.tuxcare.11.els22 for CloudLinux 8 LTS and kernel-lts-5.14.0-284.1101.el9.tuxcare.11.els22 for CloudLinux 9 LTS. To install one now, run:
yum update 'kernel*' --enablerepo=cloudlinux-updates-testing
reboot
After the reboot, uname -r should report els22 or a later build. We’ll note here when it reaches the stable channel.
Status: November 18, 2026, 14:53 UTC.
No KernelCare livepatch for this CVE has been released yet for any CloudLinux version, and there’s no date to share. KernelCare follows the vendor fixes for this issue: livepatches come once fixed kernels are released, and we’ll add each one here. When one is released, subscribed servers take it on their next update cycle, or you can apply it at once with no reboot by running kcarectl --update. CloudLinux 8 LTS has no livepatch stream; install the fixed kernel from Stream 3.
Once your stream lists a fixed kernel, reboot into it, run uname -r and compare the version with the one the stream gives. For KernelCare, run:
kcarectl --patch-info | grep CVE-2026-99288
A patched server prints lines naming the CVE. An empty result means the livepatch isn’t applied, which is the expected result today. Use --patch-info for this: kcarectl --info doesn’t list CVEs, so it makes a patched server look unpatched.
Once the fix is installed, remove the workaround you applied. Option 1 has its revert commands above; for option 2, set the queue count and rps_cpus values back to the ones you noted.
Most kernel flaws we write about need a local account, such as a hosting customer or an attacker who got in through a hacked site. NoisyTCP needs only an open TCP port to send traffic to.
CageFS and LVE don’t help here, because the kernel handles these packets before any account, cage or limit applies. A crash takes the whole server down, with every site, mailbox, and database on it, and the attacker can repeat it.
The attack is hard to pull off, though. The attacker has to win a race between two CPUs handling the same connection. The researcher measured five seconds to two minutes against virtual machines, and about fourteen hours against a quiet physical server with eight receive queues.
The researcher also describes a use-after-free as a theoretical worst case that is hard to trigger; no code execution has been shown.
To find the connection an incoming packet belongs to, the kernel searches its table of established TCP connections without taking a lock. Memory freed from one connection can be handed straight to a new one.
The lookup re-checks the connection’s addresses and ports after it takes a reference, but it doesn’t check that the connection is still in the table. An attacker makes the kernel free a pending connection and immediately build a new one at the same memory address, with the same addresses and ports. The lookup then returns the new connection, which has no listening socket attached, and the kernel reads through a null pointer and panics.
The demonstrated path goes through SYN cookies, which is why turning them off closes it. The upstream fix adds the missing check.
Nikolay Galov
Comments