enzosoler.comreviews 09bench 9 devices
Published

NoisyTCP: an 11-year-old Linux kernel bug lets anyone crash your server over TCP. Update the kernel.

A row of black mesh-door server racks with red and blue network cables inside
Image: Brett Sayles / Pexels

CVE-2026-99288, called NoisyTCP, is a race in the Linux kernel's TCP socket lookup that an unauthenticated attacker can trigger with crafted packets to panic the kernel. It affects plain TCP listeners such as Nginx, Apache and SSH on almost every distribution. Debian has fixed kernels out; Red Hat rates it Important. Update and reboot, or use the stopgaps.

If you run Linux servers that accept TCP connections from the internet, put a kernel update on this week’s list. A bug called NoisyTCP, tracked as CVE-2026-99288, lets an attacker with no login send crafted TCP packets that make the kernel crash and take the whole machine down. It sits in code added in 2015, so it reaches almost every distribution. Install your distribution’s fixed kernel and reboot. If you can’t yet, there are two stopgaps below.

What the bug does

Security researcher Hyunwoo Kim, who found the flaw and wrote the fix, describes it in his NoisyTCP write-up on GitHub. When the kernel looks up an existing TCP connection without taking a lock, it can grab a socket whose memory was just freed and reused for another connection. That leads to a NULL pointer dereference and a kernel panic. The upstream fix, merged on November 3, makes the lookup reject sockets that are no longer in the connection table and try again.

  • Who is exposed. Any server with a listening TCP port. Kim says no special configuration is needed: Nginx, Apache and SSH on their default settings all qualify.
  • The condition. The machine needs two or more network receive queues, or RPS spreading packets across CPU cores. Most multi-core servers and cloud VMs have that.
  • How fast. Kim reports a crash in 5 seconds to 2 minutes against virtual machines running those default servers. A quiet bare-metal box with 8 receive queues took about 14 hours with an exploit tuned to it.
  • Containers. In a container, the panic takes down the entire host, according to Kim.
  • What it doesn’t do. Red Hat says the flaw does not allow data changes or code execution. Kim says a second path in the same race could lead to memory corruption, but it is hard to trigger and he has only a limited lab exploit. Treat it as a remote crash for now.

Red Hat rates the bug Important, with a preliminary CVSS score of 7.5 that it says is subject to review. NVD had no score as of November 16. BleepingComputer reported the disclosure after the embargo through the linux-distros list ended.

Where the fixes stand

  • Debian. The Debian security tracker lists fixed kernels: 6.1.192-1 for Debian 12 (bookworm) and 6.12.117-1 for Debian 13 (trixie), plus 7.3.4-1 in unstable.
  • Red Hat Enterprise Linux. Red Hat’s page did not list fixed RHEL kernel versions as of November 16. Watch the CVE page for errata.
  • Ubuntu and others. We found no published fixed versions yet. Check your vendor’s security notices for CVE-2026-99288.
  • Custom kernels. Look for commit 4e2b3ef19e0e in your tree.

A crash bug matters most where one box serves many

A crash bug on one web server means downtime and a reboot. On a container host or a VM host it means every tenant goes down at once, and an attacker can repeat it as soon as the machine comes back. That makes shared hosts, Kubernetes nodes and small teams’ all-in-one servers the first to patch. A laptop or desktop that doesn’t accept inbound connections from the internet is a much smaller target.

What to do

  • Update the kernel and reboot. A new kernel package does nothing until the machine runs it. If you use live patching, confirm your vendor ships a patch for this CVE.
  • If you can’t patch yet, pick a stopgap. Kim and Red Hat give two. Setting net.ipv4.tcp_syncookies=0 blocks the race but weakens protection against SYN flood attacks, and it is set per network namespace, so apply it inside every container that has its own. The other option is to cut the receiving network interface down to a single receive queue and turn off RPS and RFS, which costs throughput.
  • Remove what you don’t need on the internet. An SSH port open to the world is one more way in. Put it behind a VPN or an allowlist.

This is the second Linux kernel TCP flaw this month. The first, CVE-2026-98432, is rated 9.8 and has its own fixed kernels, so make sure the update you install covers both. Recent server fixes worth checking in the same maintenance window: Apache HTTP Server 2.4.71 and OpenSSH on Alpine.

Related, November 16: Anthropic will disclose severe bugs it reports after 30 days instead of 90, starting with reports made in 2027.

One email when a review goes up. Unsubscribe from any of them.