NoisyTCP (CVE-2026-99288) can remotely take down a Linux server that processes TCP traffic, without credentials or any special service configuration. The flaw is in the kernel: the attack can also bring down the host and other services, even if the immediate target is an application inside a container. The primary measure is to update to the corrected kernel provided by the system vendor.
Which servers should be checked?
The risk affects Linux servers with a TCP port accessible from a location where an attacker can send traffic, if the system processes packets using multiple receive queues or distributes the workload across CPUs. Services that may be exposed include Nginx, Apache and SSH, even with default configurations. The attacker does not need to log in.
The confirmed consequence is a denial of service: the kernel may panic and halt the machine. Researcher Hyunwoo Kim also identified a second exploitation path, but said that his remote code execution test was limited and targeted laboratory virtual machines; he has not demonstrated a general-purpose code execution exploit.
Kim’s tests found that QEMU/KVM virtual machines running Nginx, Apache or SSH could crash within five seconds to two minutes. On a physical server running Nginx, it took around 14 hours. These are test results, not a guarantee that a particular machine will crash within those timeframes.
Steps to check and reduce the risk
Inventory Linux hosts that accept TCP connections from untrusted networks, including those running services in containers or virtual machines. Prioritize systems directly exposed to the internet and those hosting critical services. The flaw is in the kernel, so updating only Nginx, Apache or SSH will not fix it.
Check the active kernel version on each host using the command cited in IlSoftware.it’s advisory:
uname -rThe result identifies the kernel currently running, but is not enough on its own to determine whether the system is patched: distributions may backport the fix and retain version numbers that differ from upstream versions. Compare the result with your distribution’s security advisory and package status.
Search your system vendor’s advisories for CVE-2026-99288 and check the kernel package status for your specific distribution version. Do not assume that an old kernel is vulnerable or that one with an apparently recent version number is fixed without checking the package: backports can change how the version should be interpreted.
Install the corrected kernel using your distribution’s update mechanism. For reference, the Linux CVE team lists upstream branches 6.12.117, 6.18.63 and 7.3.4 as fixed. These versions do not replace checking the package provided by your vendor.
Branch or distribution Reported status Linux upstream 6.12 Fixed starting with 6.12.117 Linux upstream 6.18 Fixed starting with 6.18.63 Linux upstream 7.3 Fixed starting with 7.3.4 Debian bookworm The tracker marks 6.1.176-1 as vulnerable and 6.1.192-2 as fixed Debian trixie The tracker marks 6.12.107-1 as vulnerable and 6.12.117-1 as fixed Reboot the host after updating so it starts with the corrected kernel. IlSoftware.it notes that updating an application alone does not fix the operating system’s TCP stack flaw. After rebooting, check the active kernel and package status again.
If you cannot patch yet, consider the temporary mitigation published by Kim: disabling SYN cookies. The setting is:
sudo sysctl -w net.ipv4.tcp_syncookies=0The setting is managed per network namespace. If a container uses its own namespace, apply the measure in that container as well. Disabling SYN cookies weakens protection against SYN floods, so do not treat it as harmless or as a substitute for the patch.
Consider using a single RX queue only if you understand the host’s network configuration. Kim’s advisory says that reducing the number of RX queues to one on the interface receiving external traffic and disabling RPS and RFS can prevent the race condition that triggers the flaw. This may hurt performance; the advisory does not provide a universal command to apply it, so do not copy generic settings without adapting them to the system.
Review direct exposure while completing the update: identify which services accept connections from external networks and whether there is filtering in place, such as a firewall or load balancer. RedesZone recommends checking these points to limit access paths. They do not replace fixing the kernel on a vulnerable system.
“So far I have only written a very limited, lab-specific exploit that targets VM servers.”
— Hyunwoo Kim, security researcher

Mistakes that can leave you exposed
Do not rely only on the number returned by uname -r: distributions may include the fix through backports. Confirm the package status in the relevant security advisory instead of blindly comparing it with the upstream versions in the table.
Do not limit your review to containers or assume they isolate the rest of the system. The flaw is in the kernel, and Kim warns that an attack against a web server inside a container can bring down the host and other services that depend on it.
Do not confuse mitigation with the patch: disabling SYN cookies weakens protection against SYN floods and must be applied in each affected network namespace. Limiting receive queues can also affect performance. Both are temporary measures; the priority remains installing the corrected kernel and rebooting the host.
Sources: theregister, RedesZone, IlSoftware.it, GitHub, Debian Security Tracker