
This post is a translation of the Spanish original.
NoisyTCP: the patch is easy, the mitigation is the trap
Eleven years. That's how long the code that lets anyone, from outside and without credentials, trigger a kernel panic on a Linux server with an open TCP port has been in the kernel. Nginx, Apache, SSH with its default configuration. The usual.
The bug is called NoisyTCP, it's registered as CVE-2026-99288 and it was found (and fixed) by Hyunwoo Kim, who published the advisory on oss-security when the embargo with the distributions ended.
What exactly happened
It's a race condition between SYN cookie handling and the way the kernel looks up and stores TCP sockets. With crafted packets, an attacker manages to make it dereference a null pointer, and the result is a kernel panic. According to what's been published there's no code execution or data theft: it's a denial of service. But it's a denial of service that brings down the whole machine, not a process.
The condition for it to be exploitable is that the machine spreads incoming traffic across several CPUs, either with several receive (RX) queues on the card or with RPS enabled. Almost any server with a modern NIC.
The faulty code went in in October 2015 and the fix reached mainline on 3 November. The stable branches got it on the 10th in versions 6.12.117, 6.18.63 and 7.3.4. Debian already has packages for trixie and bookworm; for Ubuntu and RHEL, at the time of writing, the advisories I checked didn't yet list a fixed package. There's some dispute over severity: Red Hat scores it 7.5 and some outlets mention a 9.8 assigned by the kernel CVE team, a figure I haven't been able to confirm at the original source.
What worries me isn't the patch
Fixing it is boring: update the kernel and reboot. The interesting part is when you can't reboot yet, which is more common than it should be. For that gap, Red Hat recommends one of these two things:
- Disable SYN cookies with
sysctl -w net.ipv4.tcp_syncookies=0. - Leave exposed interfaces with a single RX queue and disable RPS and RFS.
And here comes my take, which may not be the popular one: on most servers I wouldn't touch SYN cookies. They're precisely the classic defence against a SYN flood. Removing them means trading a theoretical panic risk for a very practical one of having your connection queue saturated with something anyone knows how to launch. And the other route, serialising everything into one queue, punishes performance precisely on the busiest machines, which are the ones you most want to protect. Also, one guide suggests that route doesn't cover every entry path, although I've only seen that in one source.
There's a detail people miss: that sysctl is per network namespace. If you change it on the host and you have containers with their own network stack, each one keeps its own value. Half-mitigating gives a false sense of being covered, which is worse than knowing you aren't.
I acknowledge the best argument on the other side. A remote unauthenticated panic on an internet-facing machine is serious, and if you have a load balancer or provider in front already filtering SYN floods, disabling cookies on the server behind loses much of its cost. In that scenario, go ahead.
What I'd do
Check uname -r, see whether my distribution already has the package, and if it does, schedule the reboot today rather than "in the next window". If it doesn't, weigh case by case what is really exposed. In the homelab the risk is low, but the reboot is cheap too.
I'm left with the question I ask myself every time something like this comes out: if a bug this size slept for eleven years in one of the most scrutinised parts of the kernel, how many of the mitigations we apply in a hurry end up being worse than the bug they were meant to cover?


