How To

NoisyTCP (CVE-2026-99288): Check, Mitigate, Patch the Kernel

Patching last week’s TCP flaw didn’t close NoisyTCP. CVE-2026-99288 is a separate race in the same code path, and a kernel that carries the CVE-2026-98432 fix is still exposed to it. Any Linux server that accepts TCP from the network and spreads incoming packets over two or more RX queues, or over several CPUs with RPS, can be crashed remotely by a client that never logs in. That covers most cloud VMs on their default settings.

Original content from computingforgeeks.com - post 191433

This guide shows how to check whether your kernel carries the fix, what each distro has shipped so far, how the two published mitigations behave on a real host, and what the Debian upgrade looks like from vulnerable kernel to fixed changelog. The checks ran on GCP VMs with Debian 12 and 13, Ubuntu 24.04 and 26.04, Rocky Linux 10, AlmaLinux 9 and Fedora 45 on 16 November 2026, and the distro table reflects the trackers on that day.

Last week’s patch fixed a different race

Both bugs trace back to the same 2015 commit, 079096f103fa (“tcp/dccp: install syn_recv requests into ehash table”), which let the kernel process request sockets without holding the listener lock. CVE-2026-98432 was a use-after-free of IP options inside tcp_v4_syn_recv_sock(), and its fix only changed when those options get freed. NoisyTCP sits one step earlier, in the lockless established-socket lookup that every incoming TCP segment goes through, so the earlier fix never touched it.

The fix commit, 4e2b3ef19e0e (“inet: reject unhashed sockets during established lookups”) by Hyunwoo Kim, describes what goes wrong. In order:

  1. A lookup matches a socket’s tuple and keeps its address.
  2. Another RX context frees that socket, and the slab hands the same memory to a new SYN cookie request with the same tuple.
  3. The stale lookup takes a reference on the new object, passes the tuple recheck and returns it, even though it isn’t hashed.
  4. tcp_v4_rcv() reads through its NULL listener pointer, and the kernel panics.

Step 2 needs a second CPU working on packets for the same connection at the same moment. That’s why Kim’s oss-security disclosure limits the trigger to “servers that have two or more RX queues, or where RPS distributes packets across multiple CPUs”, and why both mitigations below go after that second CPU. The fix itself adds the missing check: a lookup that lands on an unhashed socket drops its reference and starts again.

The crash is proven, the UAF isn’t

What has actually been shown is a remote kernel panic. The fix commit describes a NULL pointer dereference in tcp_v4_rcv() that takes the whole kernel down, and Kim’s disclosure gives the timing: about 5 seconds to 2 minutes against Nginx, Apache or SSH on QEMU/KVM guests, where vCPU preemption widens the window, and about 14 hours against a quiet bare-metal Nginx host with 8 RX queues. A host running containers goes down whole, since the panic is in the shared kernel.

The disclosure also describes a second flow, a “full child socket ABA race”, that Kim says “can trigger a remote use-after-free” but is “generally hard to trigger without using special techniques”. Treat that as the reporter’s claim and nothing more. According to kernelscan’s write-up, that route came from a test kernel with artificial delays, the TCP maintainer questioned whether it can happen on a real kernel, and the paragraph was dropped from the final commit. Nobody has published a demonstration of it.

The Debian tracker links the NoisyTCP project on GitHub, and you’ll see it described as a public PoC. It isn’t one. On 16 November 2026 the repository held a single file, README.md, last committed on 15 November at 21:58 UTC, and that README is the same write-up as the mailing list post. Its FAQ answers the RCE question with “Not yet. and basically, no”, says Kim has only “a very limited, lab-specific exploit that targets VM servers”, and adds that “remote DoS can be achieved without much difficulty”. No exploit code was published there, and nothing in this guide runs one.

Scores don’t settle much yet. NVD shows no assessment. Red Hat rates it Important with a preliminary CVSS 3.1 score of 7.5 and the vector CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:H/I:H/A:H, marked “preliminary and subject to review”. That vector contradicts Red Hat’s own statement on the same page, which says “an unauthenticated remote attacker can trigger an immediate operating system crash across the network without requiring privileges” and that the flaw “does not allow data alteration or arbitrary code execution”. PR:L says privileges are needed, and C:H/I:H says data is at risk, while the text says neither. Plan around the text: an unauthenticated remote crash of the whole host.

Debian and Arch ship fixed kernels

Upstream, the fix landed in three stable releases on 10 November 2026. The kernel.org changelogs list it in 6.12.117, 6.18.63 and 7.3.4, while the latest 5.10, 5.15, 6.1 and 6.6 releases (5.10.274, 5.15.225, 6.1.192 and 6.6.162, all from 6 November) don’t carry it. Most distros are behind that. Debian has patched packages in its security archive, and Arch’s core repo already carries linux 7.3.4 and linux-lts 6.18.63, both fixed upstream releases. Treat this table as a snapshot and open the tracker before you act.

DistroStatus on 16 Nov 2026Fixed packageSource
Debian 13 trixieFixed, DSA-6628-1linux 6.12.117-1Debian tracker
Debian 12 bookwormFixed, DLA-4906-1linux 6.1.192-1Debian tracker
Debian forky / sidforky 7.2.15-1 vulnerable, sid fixedsid 7.3.4-1Debian tracker
Arch LinuxFixed upstream versions in corelinux 7.3.4.arch1-2, linux-lts 6.18.63-3Arch packages
Ubuntu 16.04 to 26.10Needs evaluation (14.04 ignored)NoneUbuntu CVE page
RHEL 8, 9 and 10Affected, no erratum (RHEL 6 and 7 not affected)NoneRed Hat CVE data
Rocky Linux 10, AlmaLinux 9No advisory in dnf updateinfo, newest kernel changelog silentNoneLab check
Fedora 45No advisory, newest kernel in updates (7.3.3) lacks the fixNone in updatesLab check

Red Hat’s rows come from its CVE data rather than its errata: the machine-readable feed for the CVE lists the kernel package of RHEL 8, 9 and 10 and kernel-rt on RHEL 8 and 9 as “Affected”, plus RHEL CoreOS in OpenShift 4 and 5. RHSA-2026:98099, the RHEL 8 kernel update issued on 11 November, doesn’t list this CVE. Fedora tracks upstream stable closely, so its fix will most likely arrive as a 7.3.4 or newer kernel, but the f45 updates repo was offering 7.3.3 when we checked.

The changelog decides, not the version

Version numbers mislead in both directions here. Debian’s 6.1.192-1 is fixed even though upstream 6.1.192 isn’t, and a RHEL 5.14 kernel will never reach an upstream version that carries the fix. The only proof is the changelog of the kernel you’re running, so start with uname -r (if you need a refresher on reading it, see how to check the Linux kernel version). Then run one check for your distro family.

On Debian, grep the changelog shipped with the running kernel’s package. A count of 1 means the fix is in the kernel you booted, and 0 means it isn’t.

zgrep -c CVE-2026-99288 /usr/share/doc/linux-image-$(uname -r)/changelog.gz

On Ubuntu, pro fix asks Canonical’s data directly and names the source packages involved. Both of our Ubuntu VMs ran the 7.0.0-1017-gcp kernel, the image default on 24.04 and 26.04 there, and got the same verdict:

sudo pro fix CVE-2026-99288

The tool printed the full kernel CVE text, then confirmed that neither the generic linux source nor the cloud flavour has a fix.

pro fix CVE-2026-99288 on Ubuntu 26.04 reporting the CVE is not resolved

“Ubuntu security engineers are investigating this issue” matches the “Needs evaluation” rows on the CVE page. Run it again once a USN appears, and it will offer to install the fixed kernel itself.

On RHEL, Rocky and AlmaLinux, ask the advisory data and then the changelog of the running kernel only. Grep for the commit subject as well as the CVE ID, because a rebuild can carry the patch before the ID is added to its changelog. Keep the full subject: our lab run grepped only “reject unhashed sockets”, which also matches an unrelated bpf fix, as the Debian 12 changelog further down shows. The dnf cheat sheet covers the other updateinfo filters.

sudo dnf -q updateinfo list --cve CVE-2026-99288
rpm -q --changelog kernel-core-$(uname -r) | grep -c -e CVE-2026-99288 -e 'reject unhashed sockets during established lookups'
sudo dnf -q check-update kernel kernel-core

Rocky Linux 10.2 printed nothing for the advisory query, 0 for the changelog, and no newer kernel than the 6.12.0-211.74.1 it was running. Empty output from updateinfo means no advisory exists yet. It doesn’t mean you’re safe.

Rocky Linux 10.2 kernel check with an empty dnf updateinfo result for CVE-2026-99288

AlmaLinux 9.8 gave the same answer. Its newest kernel, 5.14.0-687.66.2, has a changelog dated 6 November whose top entry fixes an unrelated socket bug, CVE-2026-98276. To check a kernel you haven’t installed yet, download it and read the RPM’s changelog without installing it. On EL 8 and 9, dnf download comes from dnf-plugins-core, which the GCP images ship but a minimal install may not:

mkdir -p ~/kdl && cd ~/kdl && sudo dnf -q download kernel-core
rpm -qp --changelog kernel-core-*.rpm | grep -c -e CVE-2026-99288 -e 'reject unhashed sockets during established lookups'

Fedora is clearer about the missing advisory, because dnf5 says so in words instead of printing nothing. The candidate kernel was 7.3.3, one release short of the upstream fix:

Fedora 45 newest kernel 7.3.3 changelog with no NoisyTCP fix

On the running 7.2.13 kernel, dnf updateinfo list --cve CVE-2026-99288 answered No advisory found fixing the CVE ID: "CVE-2026-99288". Watch for a 7.3.4 or newer build in updates, then confirm it with the same changelog grep. New Fedora kernels sit in the updates-testing repo for a few days before they reach updates, so a test host with that repo enabled can get the fix earlier.

RX queues decide the exposure

An unpatched kernel is only half the condition. The other half is how many CPUs receive packets, and on every VM in this lab the answer was four. Read it from the NIC first; ip -br link gives you the interface name, which was ens4 on Debian, Ubuntu and Fedora and eth0 on Rocky and Alma. The Debian 12, Debian 13 and Fedora 45 images don’t ship ethtool, so install it first with sudo apt install ethtool or sudo dnf install ethtool.

ip -br link
sudo ethtool -l ens4
grep . /sys/class/net/ens4/queues/rx-*/rps_cpus
cat /proc/sys/net/core/rps_sock_flow_entries

All seven VMs, on 4-vCPU n2-standard-4 instances, reported the same values. The relevant lines:

Current hardware settings:
RX:		n/a
TX:		n/a
Other:		n/a
Combined:	4
/sys/class/net/ens4/queues/rx-0/rps_cpus:0
/sys/class/net/ens4/queues/rx-1/rps_cpus:0
/sys/class/net/ens4/queues/rx-2/rps_cpus:0
/sys/class/net/ens4/queues/rx-3/rps_cpus:0
0

RPS is off (every rps_cpus mask is 0), but that doesn’t help, because Combined: 4 already gives the NIC four receive queues serviced by different CPUs. Either condition is enough. You’re exposed if Combined or RX is above 1, or if any rps_cpus mask is non-zero. A single-vCPU VM or a NIC with one queue and RPS off is the only shape that falls outside Kim’s stated trigger, and that’s rare on servers.

Debian 12 and 13 patch in one reboot

The GCP Debian images from 9 November still booted vulnerable kernels, 6.1.190-1 on bookworm and 6.12.116-1 on trixie, so we didn’t need to downgrade anything to show the before state. Check what you run and what apt would install:

uname -r
dpkg -l 'linux-image-*' | grep ^ii
sudo apt update
apt-cache policy linux-image-cloud-amd64

On Debian 13 the changelog count was 0, and trixie-security offered 6.12.117-1 as the candidate:

Debian 13 running kernel 6.12.116 with no CVE-2026-99288 changelog entry and 6.12.117-1 as the trixie-security candidate

Cloud images use the cloud kernel flavour, so the meta-package is linux-image-cloud-amd64. On a regular install, the dpkg -l line above shows which meta-package you have; install that one instead. Upgrading the meta-package pulls in the new versioned kernel package.

sudo apt install -y linux-image-cloud-amd64

On bookworm, apt installed linux-image-6.1.0-55-cloud-amd64 at 6.1.192-1 from bookworm-security, along with firmware-linux-free, and ran update-initramfs for it:

The following NEW packages will be installed:
  firmware-linux-free linux-image-6.1.0-55-cloud-amd64
The following packages will be upgraded:
  linux-image-cloud-amd64
1 upgraded, 2 newly installed, 0 to remove and 6 not upgraded.
...
Setting up linux-image-6.1.0-55-cloud-amd64 (6.1.192-1) ...
update-initramfs: Generating /boot/initrd.img-6.1.0-55-cloud-amd64
Setting up linux-image-cloud-amd64 (6.1.192-1) ...

Nothing is fixed until the new kernel runs, so reboot:

sudo systemctl reboot

After the reboot, read the new kernel’s changelog around the entry. Debian 13 now ran 6.12.117 and listed the fix inside the upstream stable update, which is where you’d expect it, since upstream 6.12.117 carries it:

uname -r
zgrep -c CVE-2026-99288 /usr/share/doc/linux-image-$(uname -r)/changelog.gz
zcat /usr/share/doc/linux-image-$(uname -r)/changelog.gz | grep -n -B2 -A2 -e CVE-2026-99288 -e 'reject unhashed sockets' -e 'restore refcount ordering' | head -30

The count went from 0 to 1, and the matching line sits indented among the other upstream changes:

Debian 13 booted into 6.12.117 with the CVE-2026-99288 entry in the kernel changelog

Debian 12 is the interesting one. Upstream 6.1.192 doesn’t contain the fix, so how is Debian’s 6.1.192-1 fixed? The changelog answers it. The NoisyTCP line isn’t part of the upstream stable list; it’s a top-level Debian entry next to two xen-netback fixes, signed off by Salvatore Bonaccorso on 14 November 2026. Debian applied the patch itself.

Debian 12 kernel 6.1.192-1 changelog showing the Debian backport of the NoisyTCP fix

Ignore the second match at line 28485. “bpf: reject unhashed sockets in bpf_sk_assign” is an older, unrelated fix that happens to share the words. The last grep in the shot is the one to note: neither Debian kernel mentions “restore refcount ordering”, the second patch covered in the memory ordering section below. Also note that bookworm-backports offers a 6.12.95 cloud kernel and trixie-backports a 7.2.12 one, and neither appears as fixed on the tracker, so stay on the security kernel if you use backports elsewhere.

SYN cookies off closes one route

Until your distro ships a kernel, the disclosure and Red Hat’s page give two temporary mitigations. The first is net.ipv4.tcp_syncookies=0, which “blocks the path where a request socket is reconstructed through a SYN cookie”. That’s the route Kim’s crash uses, and it’s the only path the disclosure says the switch blocks. Kernelscan goes further and says “Do not switch off SYN cookies as a workaround”: it “removes the demonstrated crash route” but “reopens the door to SYN floods”, and “the lookup bug itself stays in place”. It also reports that the TCP maintainer named Fast Open, migrated request sockets and new connection sockets as further cases of sockets held outside the table. You’d give up SYN flood protection to close one demonstrated path.

Every image in the lab had it on, and on the Ubuntu 26.04 and Rocky 10 images we traced the value to Google’s own /etc/sysctl.d/60-gce-network-security.conf. If you decide the trade is worth it on a host, write a drop-in that sorts after it:

echo 'net.ipv4.tcp_syncookies = 0' | sudo tee /etc/sysctl.d/90-noisytcp.conf
sudo sysctl --system | tail -3
cat /proc/sys/net/ipv4/tcp_syncookies

Now the part people miss. The setting is per network namespace, and a namespace created afterwards gets the kernel default, not your host value. We read it inside a fresh namespace on Debian 13:

sudo ip netns add ctr1
sudo ip netns exec ctr1 cat /proc/sys/net/ipv4/tcp_syncookies
sudo ip netns delete ctr1

The host read 0 and the new namespace read 1:

tcp_syncookies set to 0 on the host while a new network namespace still reads 1

A Docker 29.1.3 container on Ubuntu 26.04 behaved the same: sudo docker run --rm busybox cat /proc/sys/net/ipv4/tcp_syncookies printed 1 with the host at 0. A container that publishes ports through its own namespace keeps SYN cookies on, so the host setting does nothing for it. You’d have to set it per container, for example with --sysctl in Docker or in the pod’s security context on a Kubernetes node. To undo the mitigation, remove the file and set the value back:

sudo rm /etc/sysctl.d/90-noisytcp.conf
sudo sysctl -w net.ipv4.tcp_syncookies=1

The sysctl -w line echoed net.ipv4.tcp_syncookies = 1 on Debian, Ubuntu and Rocky, and a fresh read of /proc confirmed it.

Error: “sysctl: command not found”

On Debian 13, running sysctl as a normal user over SSH failed with bash: line 1: sysctl: command not found, even though the binary is installed. It lives in /usr/sbin, which isn’t on a regular user’s PATH there. Use sudo sysctl, or read the value from /proc/sys/net/ipv4/tcp_syncookies directly, which works for any user.

One RX queue serializes the race

The second mitigation removes the second CPU instead: “Reducing the RX queues of the interface where external packets come in to one and disabling RPS and RFS can block the race.” It targets the precondition rather than one path, so it covers routes the SYN cookie switch leaves open. Apply it to the interface that faces untrusted traffic, and have console access ready first if that NIC also carries your SSH session: we only tested virtio, and some physical NIC drivers reset the device when the channel count changes, which can drop the link.

sudo ethtool -L ens4 combined 1
sudo ethtool -l ens4
for f in /sys/class/net/ens4/queues/rx-*/rps_cpus; do echo 0 | sudo tee $f; done
echo 0 | sudo tee /proc/sys/net/core/rps_sock_flow_entries

On the Ubuntu 26.04 VM, the virtio NIC accepted the change without dropping the SSH session, and only rx-0 and tx-0 were left under queues/:

ethtool -L reducing a virtio NIC to one combined queue with RPS and RFS off

The usual objection is throughput, so we measured it. With iperf3 from the Debian 13 VM to the Ubuntu 26.04 VM over the internal network, 8 parallel streams moved 9.62 Gbit/s with four receive queues and 9.82 Gbit/s with one, and a single stream did 9.87 and 9.89 Gbit/s. No measurable cost, but read that carefully: both numbers sit at the roughly 10 Gbit/s egress cap of a 4-vCPU instance, and one CPU could keep up with it. On a bare-metal host with a 25 or 100 Gbit/s NIC and many connections, a single queue puts all receive interrupts on one core, and that’s where you’ll pay. Measure on your own hardware before you leave it in place.

Two limits apply. ethtool -L only changes the running driver state, so it’s gone after a reboot unless your network configuration reapplies it. And it only covers the interface you change, so a host with several public NICs needs it on each. When the patched kernel is installed, restore the queues with sudo ethtool -L ens4 combined 4 (use the maximum ethtool -l reports).

Weakly ordered CPUs need a second patch

Everything above was tested on x86_64. On CPUs with weaker memory ordering, kernelscan reports that “the first patch alone does not close the race”, and a second mainline commit, 5b5d8f351dd9 (“net: restore refcount ordering for lockless socket access”), switches the refcount operations to acquire and release forms. Kernelscan says it’s queued for 6.12, 6.18 and 7.3 and “failed to apply” to 5.10 through 6.6. It wasn’t in the 6.12.117, 6.18.63 or 7.3.4 changelogs, and the lab grep found it in neither fixed Debian kernel. The commit itself is tagged for stable and explains why: an older refcount consolidation removed full memory ordering “on weakly ordered architectures”, so accesses after taking a reference can be reordered before it. That covers arm64, POWER and RISC-V servers, so on those, expect to patch twice.

Mixed fleets make this harder to track. The Debian, Ubuntu and Rocky comparison covers how each family’s kernel cadence differs, which is the reason the same CVE lands weeks apart across them.

Exposure checks in order

Work through a host top to bottom. The first row that settles the question is your answer, and anything that reaches the last row needs one of the two mitigations until the kernel lands.

CheckCommandResult that matters
1. Fixed kernel runningzgrep -c, pro fix or rpm -q --changelog on $(uname -r)Count 1 or “resolved”: done on x86_64 (arm64, POWER and RISC-V also need the second patch)
2. Fixed kernel installed, not booteduname -r against dpkg -l or rpm -q kernel-coreReboot
3. Multiple RX queuesethtool -l <nic>Combined or RX above 1: exposed
4. RPS ongrep . /sys/class/net/<nic>/queues/rx-*/rps_cpusAny non-zero mask: exposed
5. Interim choiceethtool -L <nic> combined 1 or tcp_syncookies=0 per namespaceSingle queue covers every route; SYN cookies off covers the demonstrated one

If you also run the checks from the CVE-2026-98432 walkthrough, don’t take a clean result there as an answer here. Each fix has its own changelog line, and only the NoisyTCP one proves this race is closed.

Keep reading

UFW Firewall Commands with Examples on Ubuntu 26.04 / 24.04 Security UFW Firewall Commands with Examples on Ubuntu 26.04 / 24.04 PHP2Shell (CVE-2026-109618, CVE-2026-109619): Check and Patch PHP Php PHP2Shell (CVE-2026-109618, CVE-2026-109619): Check and Patch PHP PHP2Shell on End-of-Life PHP 8.1, 8.0 and 7.4 (WHMCS, ionCube) Php PHP2Shell on End-of-Life PHP 8.1, 8.0 and 7.4 (WHMCS, ionCube) Git CVE-2026-111760: Check and Patch the Recursive Clone RCE Git Git CVE-2026-111760: Check and Patch the Recursive Clone RCE TrueNAS SCALE CVE-2026-123033: Check Your Version and Patch Security TrueNAS SCALE CVE-2026-123033: Check Your Version and Patch Configure SELinux Context for Podman Storage on Rocky Linux 10 / AlmaLinux 10 AlmaLinux Configure SELinux Context for Podman Storage on Rocky Linux 10 / AlmaLinux 10

Leave a Comment

Press ESC to close