Monday, 16 November 2026

Linux kernel patch fixes NoisyTCP, a network bug that can panic servers

Hyunwoo Kim's NoisyTCP flaw (CVE-2026-99288) has been in Linux since 2015 and can bring down default web servers. The fix was merged on 3 November.

Key points

  • NoisyTCP (CVE-2026-99288) can crash a Linux server over the network with no login.
  • It needs two or more receive queues or RPS; default Nginx, Apache and SSH are affected.
  • The kernel fix, commit 4e2b3ef19e0e, landed on 3 November 2026.

Anyone running a public Linux server should check for a new kernel build this week. A fix merged on 3 November closes NoisyTCP, a flaw that has shipped in Linux since October 2015 and can take down a machine from across the internet. Researcher Hyunwoo Kim disclosed it as CVE-2026-99288.

No login is needed and no odd configuration. A server only has to be listening for TCP connections, which describes Nginx, Apache, SSH and most proxies out of the box. The catch for an attacker is the hardware: the network card must spread traffic over two or more receive queues, or the machine must have Receive Packet Steering enabled. Plenty of current servers meet that.

In his advisory, Kim explains the mechanism as a timing collision. Fragmented IPv4 packets and plain TCP packets are sorted onto different processor cores, so one core can pick up a half-open connection record just as another core frees it and recycles the memory for a new request. The first core then follows a pointer to a listener that does not exist and the kernel stops with a panic. On a container host that stops every container on it, not only the exposed app.

Kim also describes a second route that could corrupt memory instead of just crashing, but it needs several of these collisions to succeed back to back. His working exploit for that is narrow and built for virtual machines, and he says this is not yet a remote code execution bug.

Speed depends on where the server runs. Under QEMU/KVM his tests succeeded within seconds to two minutes. A quiet physical server with eight receive queues held out for around 14 hours with his tuned code. That makes rented cloud machines the likeliest targets.

The patch is commit 4e2b3ef19e0e, and Kim calls it the priority. Distributions will fold it into their own kernel updates, and a reboot is needed to load it. Until then he lists two workarounds. One sets the net.ipv4.tcp_syncookies kernel setting to 0, which disables SYN cookies; each container with its own network namespace needs the same command, and the server becomes easier to knock over with a SYN flood. The other limits the card to one receive queue and disables RPS and RFS, which removes the race but costs throughput.

The advisory went public once the distributions’ embargo lapsed, so vendor kernel releases are the next thing to look for.

TPRO Team

The Tech Prolonged newsroom desk. Stories under this byline cover AI, the web and general tech, and are reviewed by our editors before publishing.