Home/ Blog/ Security news/ Article
Blog · Security news

Linux Kernel CVE-2026-99288: Pre-Auth Remote Crash, Patch and Reboot

NoisyTCP (CVE-2026-99288) lets a stranger crash Linux servers over TCP, fastest on VMs.

Dark cover card with faint curved lines and a Linux TCP flaw headline

NoisyTCP, tracked as CVE-2026-99288, asks very little of an attacker: no account, no unusual setting, no rare service. A Linux server that accepts TCP connections and spreads incoming packets over more than one CPU can be knocked offline from the internet. The surprise is which machines fall first. In the researcher's tests, virtual machines on QEMU/KVM panicked within 5 seconds to 2 minutes, while a quiet bare-metal server with 8 receive queues needed about 14 hours. For a fleet built on VPS instances, that ordering is the opposite of what most people would guess.

The full technical write-up went public on November 16, 2026, once the embargo on the linux-distros list ended. Hyunwoo Kim, who reported and fixed the bug, traced it to a kernel change from October 2015; the upstream fix was committed on November 3. The result is a kernel panic, so this is a denial of service (DoS). Nobody has shown code execution, and Kim's own answer to that question in the advisory is "Not yet. and basically, no." No exploitation in the wild has been reported, and the CVE is not in CISA's Known Exploited Vulnerabilities catalog. Debian's tracker does reference the public write-up as a proof of concept, so assume the trigger conditions are now widely understood.

Which Linux servers can a stranger crash with NoisyTCP?

Linux kernels from 4.4 onward are affected until patched, provided an attacker can send packets straight to a TCP listener such as SSH, Nginx or Apache, SYN cookies are left on (the default on most distributions), and the network card spreads incoming packets across two or more CPUs, through multiple receive queues or receive packet steering (RPS).

Three practical consequences follow for anyone running a hosting fleet.

A CDN in front of the website does not take the server out of scope. The attack needs a direct path to a listener, and nearly every server has one besides the web ports: SSH. An origin that only accepts ports 80 and 443 from a proxy's address ranges is still reachable through port 22 if that is open to the world.

Virtual servers are the soft target. Many cloud and VPS images give each instance several virtio-net queues so network load lands on more than one vCPU. That is exactly the multi-queue condition the race needs, and it explains the seconds-to-minutes crash times on KVM. A small single-queue VPS is harder to hit; a busy multi-core instance is not.

In a container, the crash belongs to the host. Kim notes that a successful attack takes down the whole machine, not the one application. On a shared hosting node or a Kubernetes worker, one exposed container is enough to drop every tenant on that box at once.

How the bug works, at the level a defender needs

When a TCP packet arrives, the kernel looks up which connection it belongs to without taking a lock. The memory behind a socket can be recycled while one of those lookups still holds its old address. If that memory comes back as a brand-new SYN-cookie connection request with the same addresses and ports, the lookup accepts the newcomer even though it was never added to the connection table. A few steps later the kernel reads through an empty listener pointer and panics. The CVE record describes the fix: the IPv4 and IPv6 lookups now confirm the socket is still in the table after grabbing it, and start over if it is not.

Kim also describes a second, harder path involving a fully established connection that could produce a use-after-free, meaning memory reused after it was released. That path is why the rating is not unanimous. Red Hat rates the flaw Important, with a preliminary CVSS 3.1 score of 7.5 that includes a high integrity impact, while its summary describes the impact as availability only. Ubuntu rates the same CVE Medium. We read the gap as a hedge on the second path, not a sign that code execution is near.

Where the fixes stand on November 16

Kernel lineFixed versionStatus
Upstream stable6.12.117, 6.18.63, 7.3.4 (mainline 7.4-rc2)Fixed; no fixed 5.10, 5.15, 6.1 or 6.6 stable release listed yet
Debian 12 (bookworm)6.1.192 (DLA-4906-1)Fixed
Debian 13 (trixie)6.12.117-1 (DSA-6628-1)Fixed
Ubuntu 16.04 to 26.10None publishedNeeds evaluation, priority Medium
Red Hat Enterprise LinuxNo product status on the CVE pageRated Important; watch errata
Source: CVE-2026-99288 record, Debian security tracker, Ubuntu and Red Hat CVE pages, checked November 16, 2026.

The upstream CVE record lists fixed stable releases only for the 6.12 (6.12.117), 6.18 (6.18.63) and 7.3 (7.3.4) lines, plus mainline 7.4-rc2. The older long-term branches, 5.10, 5.15, 6.1 and 6.6, had no fixed upstream release listed. Debian has already shipped its own backports: 6.1.192 for bookworm under DLA-4906-1 and 6.12.117-1 for trixie under DSA-6628-1. Ubuntu listed every supported release from 16.04 to 26.10 as needing evaluation, so as of November 16 Ubuntu hosts had no fixed package to install.

A kernel fix only protects you after a reboot into the new kernel, or a live patch if your vendor ships one. A host that installed the update and kept running is still exposed.

If you cannot reboot into a fixed kernel today

Both stopgaps come from the researcher and Red Hat, and both cost something. Pick based on what the server faces.

  1. Check where you stand. Compare uname -r with the table above, and run sysctl net.ipv4.tcp_syncookies. A value of 1 or 2 means SYN cookies are on.

  2. Turn SYN cookies off. Run sysctl -w net.ipv4.tcp_syncookies=0 and add the same line to a file under /etc/sysctl.d/ so it survives a reboot. The setting is per network namespace, so the host value does not cover containers with their own network stack. Set it inside each one too, for example with Docker's --sysctl net.ipv4.tcp_syncookies=0. The cost is weaker protection against SYN floods, the very attack SYN cookies exist to absorb. That trade is easier on a host whose provider scrubs floods upstream than on one that takes raw internet traffic.

  3. Or collapse receive processing onto one queue. Reduce the internet-facing interface to a single queue with ethtool -L eth0 combined 1, set /sys/class/net/eth0/queues/rx-0/rps_cpus to 0, and turn off receive flow steering with sysctl -w net.core.rps_sock_flow_entries=0. Swap eth0 for your own interface name. This removes the parallelism the race needs, at the price of all inbound packet work landing on one CPU. On a busy web node that is a throughput hit you will notice.

  4. Shrink the direct path. Restricting SSH to known addresses or a VPN does not fix the bug, but it removes the most common listener a stranger can reach.

How would you know a server was hit?

NoisyTCP leaves no login, no dropped file and no new process. The only evidence is the panic itself, and only if something recorded it before the machine went down. The CVE record includes the crash Kim reproduced, which gives you the strings to search for after an unexplained reboot.

CVE-2026-99288 crash signature (from the CVE record) and where to look on the previous boot
BUG: kernel NULL pointer dereference, address: 0000000000000012
Oops: Oops: 0000 [#1] SMP NOPTI
RIP: 0010:tcp_v4_rcv+0x44e/0x1280
journalctl -k -b -1 | grep -E 'NULL pointer dereference|RIP: .*tcp_v[46]_rcv'
ls -l /var/crash/

The offset after tcp_v4_rcv changes from build to build, and an IPv6 trigger would likely surface in tcp_v6_rcv instead, so match on the function name, not the exact line. Two caveats decide whether you see anything at all. A panic often never reaches the disk, so kdump or the hypervisor's serial console log is the more reliable witness. And on distributions where the journal is not persistent, the previous boot's kernel log is simply gone; our guide to Linux logging with syslog covers keeping it. Several hosts rebooting on the same day with tcp_v4_rcv in the trace is the pattern worth an alert.

The second remote TCP bug in three weeks

The Linux kernel is the most-logged product in our own 90-day vulnerability triage ledger, at 1,806 entries. Most of them never become a post here. NoisyTCP is different because it is remote and needs no credentials. It is also the second such bug in the kernel's TCP connection code we have covered since late October, after CVE-2026-98432, the MPTCP use-after-free. Both are races in connection setup, and both need a reboot to fix.

That is the real lesson for anyone running a hosting fleet. The fix for a kernel network bug is never just a package: it is a reboot window, and hosts that reboot quarterly are now carrying both flaws. Kim says a general-purpose exploit is still being researched. If that lands, NoisyTCP stops being an uptime problem and becomes a containment one, which is a good reason to schedule the reboot this week rather than next month. For the container side of that risk, our write-up on a kernel container escape covers why one host kernel is a shared fate for every workload on it.

Frequently asked questions

What is NoisyTCP (CVE-2026-99288)?

NoisyTCP is a race condition in the Linux kernel's TCP connection lookup that lets a remote attacker crash a server without logging in. It was introduced in a 2015 kernel change, reported by researcher Hyunwoo Kim, and fixed upstream on November 3, 2026. The technical write-up went public on November 16, 2026.

Can NoisyTCP give an attacker code execution on a Linux server?

No code execution has been shown. The demonstrated impact is a kernel panic, which is a denial of service. The researcher describes a second, harder path toward a use-after-free and says a general exploit is still being researched, so treat it as a crash bug today while patching promptly.

Which kernel versions fix CVE-2026-99288?

Upstream stable kernels 6.12.117, 6.18.63 and 7.3.4 contain the fix, along with mainline 7.4-rc2. Debian ships fixed kernels 6.1.192 for bookworm and 6.12.117-1 for trixie. Ubuntu listed all supported releases as needing evaluation on November 16, 2026. The fix takes effect only after a reboot.

Does disabling SYN cookies protect against NoisyTCP?

Yes, setting net.ipv4.tcp_syncookies to 0 blocks the crash path the researcher demonstrated, but it weakens protection against SYN flood attacks. The setting is per network namespace, so containers with their own network stack need it applied separately. It is a stopgap until you reboot into a fixed kernel.

Are virtual machines and containers more exposed to CVE-2026-99288?

Virtual machines crashed fastest in testing, within 5 seconds to 2 minutes on QEMU/KVM, versus about 14 hours on a quiet bare-metal server. Containers share the host kernel, so a successful attack against one exposed container takes down the whole host and every workload on it.

Is CVE-2026-99288 being exploited in the wild?

No exploitation in the wild had been reported as of November 16, 2026, and the CVE was not in CISA's Known Exploited Vulnerabilities catalog. A detailed public write-up exists, and Debian's tracker references it as a proof of concept, so the conditions needed to trigger it are now public.

See what Suriq finds on your own servers.

Agentless checks for uptime and your cloud, a lightweight agent for detection on each server, and one console that ranks and explains what it finds.