
On November 15, security researcher Hyunwoo Kim published an advisory for a Linux kernel bug that had been sitting in the TCP stack for eleven years. The bug is called NoisyTCP, tracked as CVE-2026-99288. An unauthenticated attacker can crash a server by sending crafted packets to any open TCP port. Default Nginx, Apache, or SSH will do.
The fix landed upstream on November 3, before the CVE was even public, but most distributions still have not shipped it. That gap is the story.
What the bug actually does
Linux looks up established TCP connections without taking locks. That keeps the network stack fast, and it means socket memory can be reused for a brand new socket while an old lookup is still holding the address.
NoisyTCP works that gap. A stale lookup finds a pending connection, another receive context frees it, and the same memory comes back as a new request socket built from a SYN cookie, with the same connection tuple. The stale lookup grabs the new object, the tuple check passes, and the kernel hands back a socket with no listener attached. The TCP receive path dereferences the NULL listener pointer, and the kernel panics.
The race needs two receive contexts at once: two or more receive queues, or Receive Packet Steering spreading packets across CPUs. That is the default on most multi-vCPU servers and cloud VMs. Kim forces the collision with IPv4 fragments, which can hash to a different CPU than the rest of the flow. On QEMU/KVM virtual machines running default Nginx, Apache, or SSH, Kim crashed the host in five seconds to two minutes. On a quiet bare-metal Nginx server with eight receive queues, a tuned proof of concept took about fourteen hours.
A container does not help: the panic is in the shared kernel, so one crash takes down every container on the host. A second trigger flow could lead to a remote use-after-free, but Kim says it is hard to trigger without special techniques and nobody has demonstrated code execution. His answer on remote code execution is, "Not yet. and basically, no." Plan around a remote crash, not a remote shell.
The score does not match the statement
Red Hat rates NoisyTCP Important, with a preliminary CVSS 3.1 score of 7.5 and the vector AV:N/AC:H/PR:L/UI:N/S:U/C:H/I:H/A:H. It claims the attacker needs low privileges and can hit confidentiality and integrity, while Red Hat's own statement on the same page says neither is true. Rescored to match the statement, no privileges and availability as the only impact, CVSS 3.1 gives 5.9, Medium. A hypothetical calculation, not a published score, and the technical risk is the same either way. Oracle rates it 8.1, Canonical lists it Medium, and NVD has published the CVE record with no assessment of its own yet.
If your patch SLA ranks by CVSS number alone, a score revised downward later could push this down the queue. For internet-facing hosts, go by the description: an unauthenticated remote crash, demonstrated, against default configurations.
Who has the fix and who is still waiting
Patch status as of November 17, 2026:
| Distribution | Status | Fixed kernel |
|---|---|---|
| Upstream stable 6.12, 6.18, 7.3 | Fixed | 6.12.117, 6.18.63, 7.3.4 |
| Upstream 6.6, 6.1, 5.15, 5.10 | No fixed release yet | None listed |
| Upstream 7.2 | Branch ended before the fix | None, move to 7.3.4 |
| Debian 13 (trixie) | Fixed | 6.12.117-1 (DSA-6628-1) |
| Debian 12 (bookworm) | Fixed | 6.1.192-1 (DLA-4906-1) |
| Oracle Linux UEK | Fixed | ELSA-2026-500630, -31, -32 |
| CIQ Rocky Linux builds | Fixed Nov 15 | Patched kernels in CIQ advisory |
| RHEL 8, 9, 10 | Affected, no release date | None released |
| AlmaLinux, Rocky, Oracle RHCK | Affected | None released |
| Ubuntu 16.04-26.10 | Needs evaluation | None released |
| RHEL 6, RHEL 7, CentOS 7 | Not affected | Not needed |
Two caveats before you trust a version number. Debian backported the fix into 6.1.192-1 while upstream 6.1.192 lacks it, so check the installed package against your vendor's advisory, not the upstream version. On Debian, zgrep -c CVE-2026-99288 /usr/share/doc/linux-image-$(uname -r)/changelog*.gz returns nonzero when the booted kernel's changelog lists the fix. On arm64, POWER, and RISC-V, the three fixed stable kernels carry only half the fix: a second commit, 5b5d8f351dd9, restores memory ordering and was not part of 6.12.117, 6.18.63, or 7.3.4, so check your kernel carries both.
Check your own hosts in under a minute
Five commands answer most of the questions. Swap in your own interface name for eth0:
uname -r
ss -ltn
ethtool -l eth0
grep . /sys/class/net/eth0/queues/rx-*/rps_cpus
sysctl net.ipv4.tcp_syncookies
uname -r gives the running kernel, ss -ltn lists reachable TCP listeners, the next two lines show the race precondition (Combined or RX above 1, or any nonzero rps_cpus mask, means yes), and the sysctl line shows whether SYN cookies are on. Run them across the fleet through Ansible, Salt, or a plain SSH loop.
The two stopgaps, and what each costs
If your distribution has a fixed kernel, install it and reboot. That is the fix. If it does not, Kim's advisory and Red Hat list the same two temporary mitigations, both with a cost, so try each on a staging host first.
The first is turning off SYN cookies:
sysctl -w net.ipv4.tcp_syncookies=0
The demonstrated crash path needs a request socket rebuilt from a SYN cookie, so disabling them blocks it. The setting is per network namespace, so set it inside every container with its own namespace. The cost is weaker SYN flood defense.
The second goes after the race directly. Drop the internet-facing interface to a single receive queue and turn off RPS and Receive Flow Steering:
ethtool -L eth0 combined 1
for q in /sys/class/net/eth0/queues/rx-*/rps_cpus; do echo 0 > "$q"; done
sysctl -w net.core.rps_sock_flow_entries=0
This removes the concurrency the race depends on, not just one path. The cost is throughput, since all receive processing lands on one core, and changing channels can briefly drop the link on some NICs, so keep console access handy. The change does not survive a reboot unless the network configuration reapplies it.
If a host has no fixed kernel and can spare the throughput, the single-queue option is the more complete stopgap. If throughput cannot drop, turning off SYN cookies is the narrower choice. Remove the workarounds only after confirming the fix is active.
The durable point
Eleven years is a long time for a bug to sit in code every incoming TCP segment passes through. The commit that introduced it, 079096f103fa from October 2015, let request sockets be processed without holding the listener lock. It shipped in Linux 4.4, and it is the same commit behind October's TCP use-after-free, CVE-2026-98432. Patching that one does not fix this one. If you checked your fleet for the October flaw and stopped there, check it again.
Availability bugs get underrated because they do not steal data. A remote, unauthenticated way to take a production host offline on demand is a serious problem even at a 5.9, and the demonstrated trigger against a default VM is measured in seconds, not hours.
The patch gap is worth planning around. The fix existed upstream before the CVE went public, yet a week later RHEL and Ubuntu still have nothing to install. A patching process needs a "vendor has not shipped it yet" branch too: check the fleet, pick a stopgap, write down when to remove it, and check the vendor tracker again tomorrow.
Sources
- Hyunwoo Kim, NoisyTCP: Linux Kernel TCP Pre-auth Remote Corruption (CVE-2026-99288), oss-security, November 15, 2026
- CVE-2026-99288 Detail, National Vulnerability Database, published November 10, 2026, last modified November 15, 2026
- TuxCare, NoisyTCP CVE-2026-99288 Can Crash Linux Servers Remotely, November 16, 2026
- Linux kernel, inet: reject unhashed sockets during established lookups (commit 4e2b3ef19e0e), merged November 3, 2026



