A Linux kernel TCP bug disclosed on November 16 lets a remote, unauthenticated attacker crash a server by sending it network traffic. CVE-2026-99288, which researcher Hyunwoo Kim calls NoisyTCP, is a race condition in the kernel’s lockless lookup of established sockets. It dates to a 2015 commit that shipped in the 4.4 series, so nearly every distribution kernel from the last decade is affected. The target needs to be listening on a TCP port and receiving packets on more than one CPU. That describes most production web, SSH and API servers and the bulk of cloud VMs.
The fix went into the networking tree on November 3 and mainline on November 5, and it is in the stable releases. Distribution kernels are not all patched yet. No in-the-wild exploitation or public exploit has been reported so far. Kim has published enough detail that a working trigger is a matter of time.
Technical Details
The kernel looks up established sockets without taking locks; it uses RCU. That lookup can race with a request socket (the minisock created during the handshake) being freed and reallocated. A stale lookup on one CPU and the free/reuse of the same request socket on another CPU have to run at the same moment, in different RX contexts. When they do, the kernel ends up working on a socket that is no longer hashed. The usual result is a NULL pointer dereference and a kernel panic.
Kim describes a second path that leads to a use-after-free. It is much harder to trigger, and neither Kim nor any vendor claims code execution. Red Hat says the flaw does not allow data modification or code execution and scores it CVSS 7.5 (availability only). The kernel CNA published the CVE without a score.
What the attacker needs:
- Two or more RX queues, or RPS/RFS spreading packet processing across CPUs. This is the default on almost any multi-vCPU VM with a virtio-net or ENA/gVNIC multiqueue NIC.
- A reachable TCP listener: nginx, Apache, sshd, anything else.
- The SYN cookie path. The demonstrated trigger uses the fact that IPv4 fragments and regular TCP segments get different receive hashes. That lets the attacker put two packets for the same flow on different CPUs and line up the race.
No authentication or completed handshake is needed. Researchers reproducing it say a vulnerable cloud VM goes down within seconds once the race is being driven.
Affected and Fixed Versions
| Branch | Fixed in |
|---|---|
| 6.12 LTS | 6.12.117 |
| 6.18 LTS | 6.18.63 |
| 7.3 stable | 7.3.4 |
| Mainline | commit 4e2b3ef19e0e (“inet: reject unhashed sockets during established lookups”) |
Weakly ordered architectures need a second patch. On arm64, POWER and RISC-V, the stable releases above carry only half the fix. A follow-up memory-ordering patch, 5b5d8f351dd9, is also required. If you run Graviton, Ampere or other arm64 fleets, confirm your vendor kernel has both commits. Don’t trust the version number alone.
Older LTS branches (6.6, 6.1, 5.15, 5.10) had no backport at the time of disclosure, according to kernel tracking services.
Distribution status as of this writing:
- Debian: fixed for bookworm and trixie.
- Oracle Linux UEK: fixes shipped.
- Red Hat: RHEL 8, 9 and 10 affected; no fixed kernel yet.
- Ubuntu: still evaluating.
Impact
This is a reboot-on-demand bug, not a breach. At scale that still hurts: one attacker can take down every exposed node behind a load balancer, and they can do it repeatedly. Health checks, autoscaling and Kubernetes node replacement will absorb some of it, but they won’t fix a fleet that panics again as soon as it is back. Shared-kernel environments are hit hardest. Containers on a node all go down together, because the panic is in the host kernel.
Mitigation
- Patch. Move to 6.12.117, 6.18.63 or 7.3.4 or later, or to your distro’s fixed kernel when it ships. On arm64, POWER and RISC-V, also confirm
5b5d8f351dd9is present. Livepatch customers should check with their vendor. - Interim workaround (researcher- and Red Hat-recommended): limit the exposed interface to a single RX queue and disable RPS/RFS:This costs throughput on busy hosts. Test it before rolling it out.
1 2 3ethtool -L eth0 combined 1 for q in /sys/class/net/eth0/queues/rx-*; do echo 0 > $q/rps_cpus; echo 0 > $q/rps_flow_cnt; done sysctl -w net.core.rps_sock_flow_entries=0 - Alternative:
sysctl -w net.ipv4.tcp_syncookies=0closes the demonstrated trigger path. It also removes your SYN flood protection. The setting is per network namespace, so it has to be applied inside each container netns as well. Treat this as a stopgap for hosts protected by an upstream DDoS layer. - Reduce exposure. Put admin listeners (SSH, kubelet, metrics) behind a VPN or allowlist, and dropping IPv4 fragments at the edge where your traffic doesn’t need them makes the published trigger harder.
- Detect. Look for unexplained panics with stack traces in the inet/TCP established-lookup path (
__inet_lookup_established,tcp_v4_rcv) in kdump or console logs. Correlate them with bursts of fragmented inbound traffic.