Skip to main content

NoisyTCP Linux Kernel Flaw Lets Attackers Crash Servers Remotely Without a Login

CVE-2026-99288, dubbed NoisyTCP, lets an attacker crash most Linux servers over TCP with no login. Virtual machines and container hosts fall fastest. Where the fixes stand and what to patch first.

Detailed view of fiber optic cables connected to a server rack, showcasing modern technology.
KBY News Desk7 min read

In this article

Share

A flaw in the Linux kernel’s TCP code lets an attacker crash a server from across the network, with no login and no special configuration on the target. It is tracked as CVE-2026-99288 and its finder, the researcher Hyunwoo Kim, calls it NoisyTCP. Kim published the details on the oss-security mailing list on Sunday 15 November, once the embargo agreed with Linux distributors had ended. The fix went into the upstream kernel on 3 November. Red Hat rates the flaw Important, with a CVSS score of 7.5. On Monday evening, UK time, Debian had shipped fixed kernels and Ubuntu was still listing every supported release as “Needs evaluation”. None of the advisories reports it being used in attacks.

The short version for IT teams: this is a denial of service, not a break-in, but it can take down a whole host, and it is easiest to trigger on virtual machines. That makes cloud servers and container hosts the place to start.

#What Kim found

According to Kim’s advisory, the bug was introduced in a kernel change dated 3 October 2015 and fixed by commit 4e2b3ef19e0e on 3 November 2026. That puts it in the kernel for about 11 years, and Kim says it “affects almost all Linux distributions”.

The advisory says the flaw can be triggered against “plain” TCP listeners, which is why Nginx, Apache and SSH servers running their default configuration are named as examples. The only condition on the target is that it has two or more receive (RX) queues on its network interface, or that Receive Packet Steering (RPS) spreads incoming packets across several CPUs. We would expect most multi-core servers, physical or virtual, to meet one of those. The attacker needs to reach an open TCP port directly.

Kim describes two ways in. The first ends in a kernel panic: a remote denial of service. The second can cause a use-after-free, but the advisory says it is “generally hard to trigger without using special techniques”. Kim says he has written only “a very limited, lab-specific exploit that targets VM servers” and has not shown remote code execution. Red Hat’s own statement says the flaw “does not allow data alteration or arbitrary code execution”.

#How the crash happens, in plain terms

When a TCP packet arrives, the kernel looks up which connection it belongs to. That lookup is done without a lock, for speed, and the memory behind a connection can be freed and reused while another CPU is still looking at it. The kernel’s own fix description, published with the CVE, explains that the lookup checked the connection details a second time but did not check whether the object it had found was still a live, listed connection. If the old memory had just been reused for a new half-open connection with the same addresses and ports, the lookup could pick up the new object, which has no listening socket attached, and the kernel would follow an empty pointer and panic.

The fix adds that missing check to both the IPv4 and IPv6 lookups and makes the kernel restart the lookup when it fails. Our reading: this is a classic race between two CPUs handling packets at the same moment, which is why the number of receive queues matters and why timing decides how quickly it works.

#Why virtual machines are the soft target

The timings in the advisory are the most useful part for anyone deciding what to patch first. Against Nginx, Apache and SSH servers with default settings on QEMU/KVM virtual machines, Kim says the crash took “about 5 seconds to 2 minutes”. Against Nginx on a quiet bare-metal server with eight RX queues, it took “about 14 hours”, even with an exploit tuned to that machine.

The advisory puts the difference down to the hypervisor: on QEMU/KVM, a virtual CPU can be preempted, which in our reading stretches the narrow window the race needs. Whatever your hypervisor, we would treat an internet-facing Linux VM as crashable in minutes until it is patched.

Containers raise the stakes. Because the bug is in the kernel, Kim notes that crashing a web server inside a container “takes down not only the web server application but the entire host”. On a Kubernetes node that means every pod on it, not just the one an attacker could reach.

#Who is and is not affected

Affected: Linux systems running a kernel without the 3 November fix that accept TCP connections from untrusted networks and have more than one receive queue or RPS enabled. That covers most internet-facing Linux servers, VMs, container hosts and Linux-based appliances such as reverse proxies.

Less exposed: servers that only accept TCP from inside a trusted network, and, in our reading of the advisory’s “reach directly” condition, back-end servers that sit behind a load balancer or proxy that ends the TCP connection itself. The proxy in front of them, if it runs Linux, is exposed instead.

#Where the fixes stand

Debian’s security tracker lists fixed kernels for bookworm (6.1.192-1, DLA-4906-1) and trixie (6.12.117-1, DSA-6628-1). Ubuntu’s CVE page showed every listed release, including 24.04 LTS and 26.04 LTS, as “Needs evaluation” when we read it on Monday evening. Red Hat’s CVE page gives its Important rating and the two mitigations below. Check your own distribution’s page before assuming a kernel update you have already installed contains the fix: the upstream date alone does not tell you.

#What we would do this week

  1. List internet-facing Linux hosts, starting with cloud VMs and Kubernetes or Docker hosts, then reverse proxies and load balancers built on Linux.
  2. For each, check the running kernel with uname -r and compare it with your distribution’s advisory for CVE-2026-99288. Installing a kernel package is not enough: the host must be rebooted into it, or covered by your vendor’s live patching if it ships this fix.
  3. Patch and reboot the exposed VMs and container hosts first. A quiet physical server behind a firewall can wait for the normal cycle.
  4. Where a fixed kernel is not yet available, choose a mitigation with care. Kim and Red Hat give two. Setting net.ipv4.tcp_syncookies=0 closes the path the crash uses, but it must be set in every container network namespace as well as on the host, and it weakens protection against SYN floods. Cutting the interface to a single receive queue and turning off RPS and RFS blocks the race, at a cost in throughput.
  5. Watch for unexplained kernel panics on Linux hosts and keep the crash logs. A panic that names tcp_v4_rcv is worth a closer look.

#Leave SYN cookies on unless something else stops floods

We would not switch off SYN cookies on an internet-facing host unless something upstream, such as a cloud provider’s DDoS protection or a load balancer, already absorbs SYN floods. Otherwise you swap a race that needs a determined attacker for a flood attack that needs very little skill. For most teams the better answer is to patch the most exposed hosts quickly and accept a short window on the rest. If your distribution has not shipped a fix and a host is both internet-facing and virtualised, the single-queue workaround is the safer of the two, provided the host has throughput to spare.

#Not remote root, but more than a crash

This is not a repeat of a remote root bug, and anyone selling it as one is overstating it. But it is more serious than the “only a crash” label suggests. A remote crash that needs no login, works against default configurations and can take a container host down with every workload on it is exactly the kind of flaw that gets folded into extortion and nuisance attacks once a working tool circulates. Kim has not published one, but the advisory and the kernel’s own fix description explain the mechanism in detail.

It is also the second serious Linux kernel bug from the same researcher this year. The Register notes that Kim found Dirty Frag in May, a local privilege escalation where a working exploit went public before patches were ready. This time the process worked: the fix was upstream twelve days before disclosure and Debian had fixed packages listed by Monday. We expect the other major distributions to follow within days. If Ubuntu is still at “Needs evaluation” by the end of the week, that would be the story worth chasing.

Evidence trail

Sources and verification

Primary documentation and external technical references used in this article.

  1. 01oss-security: NoisyTCP advisory by Hyunwoo Kimopenwall.com
  2. 02NoisyTCP advisory and FAQ (GitHub)raw.githubusercontent.com
  3. 03Red Hat: CVE-2026-99288access.redhat.com
  4. 04Debian security tracker: CVE-2026-99288security-tracker.debian.org
  5. 05Ubuntu: CVE-2026-99288ubuntu.com
  6. 06The Register: NoisyTCP bug gives Linux servers a remote crash problem 11 years in the makingtheregister.com

About this page

Credited to

Dates

Published

Corrections

No corrections are listed for this page. Some review sweeps corrected pages without naming each one; those are in the corrections log.

Reader Interaction

Comments

Add a thoughtful note on NoisyTCP Linux Kernel Flaw Lets Attackers Crash Servers Remotely Without a Login. Comments are checked for spam and held for moderation before appearing.

Loading comments...

More news

All news

Was this useful?

Reports go to the site owner and are used to correct pages. Corrections are listed at /corrections.

Get new fixes by email.

New runbooks, configuration traps and fixes from KBY Technologies, about twice a week.