An 11-year-old Linux networking bug dubbed NoisyTCP lets attackers crash servers without logging in

Science & Technology

NoisyTCP: 11-year-old Linux kernel bug crashes servers

CVE-2026-99288 lets a remote attacker crash a Linux host from any open TCP port. Kim's tests: 5 seconds to 2 minutes on VMs, 14 hours on bare metal.

By
Nathan Brooks
Filed
Channel
Science & Technology
Read
4 min read

Linux stable kernels 6.12.117, 6.18.63 and 7.3.4 now ship the fix for CVE-2026-99288, a networking flaw that has lived in the kernel since a commit dated Oct. 3, 2015 and lets a remote attacker crash a machine by sending crafted TCP traffic to any open port. No login required.

Security researcher Hyunwoo Kim published the details under the name NoisyTCP, and the kernel CVE team recorded the flaw on Nov. 10. The trigger reaches plain TCP listeners, so a stock Nginx, Apache or SSH daemon is exposed even when the server "does nothing more than call listen()." Because the crash happens in kernel space, a container offers no shelter. Take down the web server inside the container and the host follows.

How does it work without a login?

The kernel looks up established TCP connections without taking locks, and the memory holding those records can be recycled while a lookup is still running. According to the CVE record, the lookup functions re-checked a connection's address details but never confirmed the record remained registered in the kernel's connection table. A lookup can therefore pick up a freshly reused record that was never registered at all.

Kim's main trigger uses SYN cookies, the standard defense against connection floods that most distributions enable by default. The recycled record ends up with no parent listening socket, and the kernel crashes when it tries to read it. A second path corrupts memory through a use-after-free, but Kim says it requires nested races and is "generally hard to trigger."

The race only works when a machine processes incoming packets on more than one CPU core, through multiple receive queues or Receive Packet Steering, which spreads packets across cores. Kim exploits the difference between how IPv4 fragments and ordinary TCP packets hash to those cores. On most modern servers, that is the default arrangement.

How fast does it crash?

On QEMU/KVM virtual machines, Kim says default Nginx, Apache and SSH servers took "about 5 seconds to 2 minutes to trigger." A quiet bare-metal server with eight receive queues running Nginx took roughly 14 hours, using a proof of concept tuned to that machine. iTnews attributes the gap to timing: a hypervisor can pause a virtual processor mid-operation, widening the window for the race.

On code execution, Kim's write-up says: "Not yet. and basically, no." It adds: "But remote DoS can be achieved without much difficulty." Red Hat likewise says the flaw permits no data alteration and no code execution. Kim has withheld his exploit code and describes it as "very limited" and lab-specific.

Which distributions have patched it?

Distribution coverage is uneven, and the numbers say so.

  • Linux stable: fix landed in 6.12.117, 6.18.63 and 7.3.4.
  • Red Hat: rates the flaw Important with a preliminary CVSS score of 7.5, but lists no fixed packages yet.
  • Ubuntu: Medium priority; 24.04 LTS and 26.04 LTS and other supported releases all listed as needing evaluation.
  • Debian: fix present in the unstable branch and in security repositories for bookworm and trixie, with forky still vulnerable.
  • Oracle: shipped a fix for its Unbreakable Enterprise Kernel on Oracle Linux 7 and 8 on Nov. 12, Cyber Kendra reported.

What can operators do before the patch lands?

Kim wrote the fix, a patch titled "inet: reject unhashed sockets during established lookups," which makes the kernel drop and retry a lookup that lands on an unregistered record. Until a patched kernel is installed and the server rebooted, Kim and Red Hat suggest two workarounds, and both carry a cost.

Turning off SYN cookies (net.ipv4.tcp_syncookies=0) must happen in every network namespace, including each container's, and it weakens protection against SYN floods. Reducing the internet-facing network card to a single receive queue and disabling packet steering blocks the race but cuts throughput. iTnews notes that the SYN cookie switch does not cover the harder second memory-corruption path.

No vendor has reported attacks using the flaw, and no public exploit code has surfaced. Kim earlier this year reported Dirty Frag, a Linux local privilege-escalation bug, iTnews reported.

With no attack in the wild and vendor patching still uneven, near-term exposure rests on operators pulling the stable kernel update or accepting the throughput hit of a single-queue workaround, and Red Hat's still-empty package list suggests the main cleanup has yet to begin.

Source: mlq.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,009 articles

Related articles

« Previous articleNext article »