Admin terminal next to a server with several network cards

NoisyTCP (CVE-2026-99288) is a flaw in the Linux kernel TCP stack, present since 2015, that lets an unauthenticated remote attacker trigger a kernel panic on servers listening on TCP. The fix landed in the kernel on November 3, 2026, and the real solution is to update and reboot. Until each distribution ships packages, though, there is time to find out which machines are exposed and apply temporary mitigations.

NoisyTCP for sysadmins in 30 seconds

  • It affects TCP services in their default configuration: web servers, SSH, proxies.
  • It is only exploitable when the server has two or more receive queues or RPS enabled.
  • Confirmed impact: denial of service; on container hosts, the whole host goes down.
  • Fix: updated kernel and reboot; temporary mitigation by disabling SYN cookies or using a single queue without RPS.

As OpenSecurity explains in its analysis of NoisyTCP (in Spanish), the bug is a race between two CPUs in SYN cookie handling that leaves a request without its listening socket, so the kernel ends up dereferencing a NULL pointer. It was found and fixed by Hyunwoo Kim (@v4bel). The confirmed impact is a system crash; remote code execution has not been demonstrated.

How to tell if a server meets the conditions

The flaw needs incoming traffic to be processed in parallel on several CPUs. That happens when the NIC uses multiple receive queues (normal on servers and on many VMs with multi-queue virtio) or when Receive Packet Steering is enabled. To check, replacing eth0 with your interface:

# Running kernel version
uname -r

# Configured receive queues (combined or rx)
ethtool -l eth0

# Queues visible to the system
ls -d /sys/class/net/eth0/queues/rx-*

# RPS mask per queue (all zeros = RPS disabled)
cat /sys/class/net/eth0/queues/rx-*/rps_cpus

# Current SYN cookie setting (1 = enabled, the default)
sysctl net.ipv4.tcp_syncookies

If there is more than one receive queue or any non-zero RPS mask, and the kernel does not yet include the fix (commit 4e2b3ef19e0e), the server meets the conditions. The most reliable way to know whether your distribution’s kernel is fixed is to check the vendor’s security advisory for CVE-2026-99288.

Temporary mitigations and their cost

The first option the researcher suggests is disabling SYN cookies. The setting is per network namespace, so on container hosts it must be repeated inside each namespace:

# Apply at runtime
sudo sysctl -w net.ipv4.tcp_syncookies=0

# Persist across reboots
echo "net.ipv4.tcp_syncookies = 0" | sudo tee /etc/sysctl.d/99-noisytcp.conf
sudo sysctl --systemCode language: PHP (php)

The cost is that the server loses its SYN flood defence. Behind a load balancer or a DDoS protection service the risk is lower; on directly exposed servers it needs weighing, and either way the change should be reverted once the kernel is patched.

The second option is to keep a single receive queue and disable RPS and RFS, which removes the race. For example, with drivers that support combined channels:

# Reduce to a single combined queue
sudo ethtool -L eth0 combined 1

# Disable RPS on existing queues
for f in /sys/class/net/eth0/queues/rx-*/rps_cpus; do echo 0 | sudo tee "$f"; done

# Disable RFS
sudo sysctl -w net.core.rps_sock_flow_entries=0

This route has a clear performance cost on busy servers, because all receive processing goes through one CPU, and ethtool changes do not survive a reboot unless added to the network configuration. For most environments the sensible plan is to prioritise patching internet-facing servers, load balancers and proxies, and container hosts, and meanwhile apply whichever mitigation best fits each service.

Frequently asked questions

How do I know if my server is vulnerable to NoisyTCP?

If its kernel is unpatched and it has two or more receive queues or RPS enabled. Check with ethtool -l and the rps_cpus files under /sys/class/net.

Is disabling SYN cookies safe?

It mitigates NoisyTCP but leaves the server more exposed to SYN floods. It is a temporary measure to revert after patching.

Are containers affected?

Yes. A crash triggered through a container takes down the whole host, and the sysctl mitigation must be applied in each network namespace.

What is the permanent fix?

Install your distribution’s updated kernel with the November 2026 fix and reboot.

Sources:

Scroll to Top