For most of the last decade, server operators assumed that a Linux machine with a current patch level and a firewall was safe from remote denial of service unless something in userspace broke. The kernel TCP stack was treated as the boring, well tested layer, hammered by every scanner on the internet and long past producing surprises. That assumption took a direct hit this week, when a security researcher published a race condition that has sat quietly in the networking code since October 2015.
The flaw is tracked as CVE-2026-99288 and named NoisyTCP by its finder, Hyunwoo Kim, who publishes as V4bel. It is a race in the way the kernel looks up established TCP connections, and the consequence is blunt: a single reachable listening port on an unpatched host lets an unauthenticated attacker drive the machine into a kernel panic. No credentials, crafted application payload, special socket option or netfilter rule is required. The attacker only needs to reach a service that accepts plain TCP connections, which describes essentially every web and SSH server on the public internet.
The disclosure arrived on the oss-security mailing list after an embargo coordinated with the linux-distros maintainers expired. Kim's report is unusually detailed for a denial of service bug, and it is paired with a patch merged on November 3, 2026, thirteen days before the public posting. The remediation path therefore already exists, but the October patch round that many administrators rely on does not contain it, and hosts considered fully up to date at the start of the month are still exposed on default installations of Nginx, Apache and OpenSSH.
Key Facts
Kim's advisory, posted to oss-security on November 16, 2026, describes a pre-authentication TCP race that panics the kernel by dereferencing a NULL listener pointer. A stale lookup in the ehash table finds an existing request socket just as another receive context frees it and reallocates a SYN cookie request with the same tuple at the same slab address. The stale lookup takes a reference and passes the tuple recheck, but the replacement has no listener attached, so the receive path follows a NULL pointer and halts the machine. iTnews reported on November 16 that Red Hat rates the flaw 7.5 out of 10 and classifies it as Important.
The trigger requirement decides who is exposed. Kim writes that packet processing must be spread across CPUs, which he achieves using the difference in receive hash between IPv4 fragments and regular TCP packets. Servers with two or more receive queues, or hosts where Receive Packet Steering distributes packets across cores, meet that condition by default. A single queue machine with RPS disabled has nothing to race, which is why the same kernel can be trivially crashable on one host and effectively immune on another.
Timing separates a theoretical bug from an operational emergency. Kim measured five seconds to two minutes against default Nginx, Apache and SSH servers on QEMU/KVM virtual machines, where the hypervisor can preempt a virtual CPU mid sequence and widen the window. On a quiet bare metal server with eight receive queues running default Nginx, the same proof of concept needed about fourteen hours. ThreatAft reported on November 16 that no attacks in the wild have been recorded, that EPSS sits at 0.16 percent and that no public exploit code has been released.
The affected range is broad and the patch matrix is awkward. The bug entered commit 079096f103fa on October 3, 2015 and was fixed in 4e2b3ef19e0e on November 3, 2026, which puts the latent period at about eleven years and the affected base at Linux 4.4 and later. Fixes reached stable as 6.12.117, 6.18.63 and 7.3.4, all on November 10, 2026, with mainline 7.4-rc2 also carrying the change. FossLinux reported on November 15 that the 7.2 branch closed at 7.2.15 on November 1, two days before the fix was written, so no 7.2 release will ever carry it, and that the 5.10, 5.15, 6.1 and 6.6 branches had no fixed upstream release at the time.
Distribution status on November 15 was uneven. Ubuntu listed every supported release from 16.04 through 26.10 at priority Medium with no fixed package, each marked as needing evaluation. Debian had shipped bookworm in 6.1.192-1 under DLA-4906-1 and trixie in 6.12.117-1 under DSA-6628-1, while Oracle covered its kernels with ELSA-2026-500630 through ELSA-2026-500632, rated 8.1 Important.
Analysis
What this really means is that the exposure here is not a fringe configuration or a hardening failure. ThreatAft reported on November 16 that the default configuration on modern server hardware is the configuration that exposes the flaw, and that assessment is hard to argue with. The kernel performs established socket lookups without a lock, socket memory is backed by SLAB_TYPESAFE_BY_RCU so a freed socket's address can be recycled while an RCU reader still holds it, and the lookup rechecks the tuple without checking whether the replacement socket is actually hashed. Each choice buys performance operators want, and together they widen a narrow window as machines gain cores.
This is worse than a typical denial of service because of the container multiplier. Kim notes that against a web server running in a container the panic takes down not only the application but the entire host. FossLinux made the same point from field experience on November 15, observing that most affected hosts run containers and that one crafted connection therefore ends every tenant on the machine. A bug that would otherwise be a single service outage becomes a multi tenant incident traced to a single packet source.
The second path deserves measured language rather than alarm. Kim describes a full child socket ABA race that can produce a remote use-after-free, but it requires nested races and is generally hard to trigger without special techniques. He does not claim remote code execution, and his own FAQ answers the RCE question with a flat no. The honest reading is that this is a reliable remote crash and, in laboratory conditions, a potential memory corruption primitive, not a wormable exploit. That distinction should shape how fast teams respond, not whether they respond at all.
The bigger picture here is that eleven year old memory safety bugs in the network stack do not get less dangerous with age, they get more widely deployed. This defect predates most of the container orchestration layers now running on top of it, which is why a 2015 race is showing up as a 2026 cloud problem. The patch gap compounds it. October's kernel updates did not include the fix, the 7.2 branch closed before the fix existed, and several long term branches had no fixed release at disclosure. Administrators who track whether they are current, rather than whether their kernel contains commit 4e2b3ef19e0e, will draw the wrong conclusion.
Why It Matters
The blast radius is set by the listening port, not by the application. Nginx, Apache and SSH in default configuration are all reachable targets, and the flaw needs no authentication, no malformed HTTP request and no SSH handshake. Any host with two or more receive queues or RPS enabled is a candidate, which covers ordinary cloud instances, and five seconds to two minutes on QEMU/KVM guests means the practical barrier to an attacker is a shell loop rather than an exploit development project.
The cascading failure mode is the part infrastructure teams should internalize. ThreatAft notes that request sockets build up until servers stop responding, that automatic traffic management then shifts users to healthy servers, and that the same defect is waiting there. A load balancer doing exactly what it was configured to do becomes an amplifier, turning one saturated host into a rotating outage across a pool. The failure will look like a mysterious sequence of reboots under load, the pattern most easily misattributed to hardware or to the application.
Next Up
The immediate action is to confirm the running kernel against the fixed versions rather than against a distribution patch date. Stable 6.12.117, 6.18.63 and 7.3.4 carry the fix, mainline 7.4-rc2 does as well, and Debian and Oracle have published their own errata. Where a fixed kernel is not yet available, the interim choices are to drop to a single receive queue with RPS and RFS off, which blocks the race, or to disable SYN cookies per network namespace, which is weaker and does not cover the second path. Neither is a substitute for the upgrade.
Expect the picture to move quickly. Ubuntu's needs evaluation status across 16.04 through 26.10 will not survive contact with a public advisory, and the long term branches at 5.10, 5.15, 6.1 and 6.6 that lacked upstream fixes at disclosure are the ones to watch. The 7.2 branch is a dead end and its users need a plan that does not involve waiting. Kim has published the full technical report, so defenders and attackers now work from the same document, and the only variable left is how fast the reboot happens.
Comments (0)
Log in or sign up to leave a comment.
No comments yet. Be the first to share your thoughts.