Dev.to Security 🔐 Cybersecurity 👁 0 📖 6 min read

Five Seconds on a VM, Fourteen Hours on Bare Metal: What NoisyTCP Actually Teaches

The same Linux kernel bug, with the same proof of concept, crashed a default Nginx server in as little as five seconds in one environment and needed about fourteen hours in another. That spread, measured by the researche

The same Linux kernel bug, with the same proof of concept, crashed a default Nginx server in as little as five seconds in one environment and needed about fourteen hours in another. That spread, measured by the researcher who found the bug, is the most instructive number in the NoisyTCP disclosure, because it exposes something security teams rarely price correctly: how much the environment decides whether a theoretical vulnerability is a practical one.

NoisyTCP, tracked as CVE-2026-99288, is a race condition in the Linux kernel's TCP networking code. Researcher Hyunwoo Kim reported it to the kernel and distribution maintainers, the upstream fix landed on November 3, and the full advisory appeared on the oss-security mailing list on November 15 after the agreed embargo expired. The flawed code was introduced in October 2015, which means the bug sat in the kernel for about eleven years and affects almost every Linux distribution.

What the bug actually is, in plain terms

When a TCP connection arrives, the kernel looks up connection state without taking a lock, because locking on every packet would be slow. Kim found an ABA race in that lockless lookup: one receive context finds a connection object, and before it can safely use it, a second context frees that memory and reallocates the same slab address for a new connection request built from a SYN cookie, with the same connection tuple. The first context's checks still pass, because the tuple matches, but the new object has no listener attached to it. The kernel then dereferences that empty pointer and panics.

Two consequences matter. First, this is a pre-authentication, remote trigger: anyone whose TCP packets can reach a listening port can attempt it, and it works against Nginx, Apache, and SSH running with default settings. No exotic configuration is required; the server only needs two or more network receive queues, or Receive Packet Steering enabled, so that two packets are processed on different CPUs at the same moment. Second, because the crash is in the kernel, a vulnerable web server inside a container does not lose just the container: the whole host goes down.

There is also a second, harder trigger path that Kim describes as a remote use-after-free, requiring several nested races to succeed. He calls it generally hard to trigger without special techniques.

The timing spectrum is the story

Against default Nginx, Apache, and SSH on QEMU/KVM virtual machines, Kim's proof of concept crashed targets in about five seconds to two minutes. Against a quiet bare-metal server with eight receive queues, a proof of concept tuned for that specific environment took about fourteen hours.

The difference is timing mechanics, not effort. The race needs another receive context to slip into a narrow window between two operations. Hypervisors can preempt a virtual CPU mid-operation, stretching that window wide open. A quiet physical machine with little scheduling variance keeps the window narrow, so the attacker has to knock for hours.

This cuts both ways for defenders. The reassuring reading is that noisy, contended, well-provisioned bare metal is harder to crash. The uncomfortable reading is that the default posture of modern infrastructure, virtual machines on shared hypervisors, is precisely the posture where this bug is fastest. "Exploitable in theory" and "crashed within the minute" were separated by a deployment detail most asset inventories do not record.

Is it RCE? The honest answer is layered

Asked whether NoisyTCP is a remote code execution bug, Kim's answer is: "Not yet. and basically, no." He has written only a limited, lab-specific exploit targeting virtual machines, is still researching a general-purpose one, and says remote denial of service is achievable without much difficulty. Red Hat's own statement says the flaw causes a kernel panic but "does not allow data alteration or arbitrary code execution."

And here a small lesson in reading advisories: the same Red Hat page, as fetched on November 16, displays a preliminary CVSS vector with high confidentiality and integrity impact ratings, and its weakness section notes that in rare circumstances NULL pointer writes can lead to code execution, while the statement text rules code execution out. These are different claims sitting on one page. The defensible reading today is the narrow one: a demonstrated remote crash, a second memory-corruption path that is hard to trigger, and a researcher who explicitly reserves judgment on what further work might achieve. No public exploit code has been released; Kim's repository contains the write-up, not a weapon.

The patch exists. Your kernel may not have it

The upstream fix, titled "inet: reject unhashed sockets during established lookups", is in stable kernels 6.12.117, 6.18.63, and 7.3.4. Distribution coverage as of November 15, per an independent status survey by Brian Vitko: Debian bookworm and trixie shipped fixed kernels; all supported Ubuntu releases were marked "needs triage" with no fixed kernel yet; RHEL 8, 9, and 10 were listed as affected with no fixed kernel yet, while RHEL 6 and 7 are not affected.

That asymmetry is the operational core of this week. The vulnerability being "fixed" upstream and your fleet being protected are events separated by distribution engineering, backports, and your own reboot schedule. Vitko's practical advice deserves repeating: check the distribution's kernel changelog for the CVE number or the patch title, not the version string, because distributions backport fixes into kernels whose versions look old, and a kernel built before the CVE number existed will list the fix only by its title.

Mitigations are relocations, not solutions

For systems that cannot take the fixed kernel yet, Kim documented two stopgaps, and each one moves risk rather than removing it.

Disabling SYN cookies closes the path where the malicious request is reconstructed from a cookie. The setting is per network namespace, so every container with its own namespace needs it too, and turning it off weakens protection against classic SYN floods. It also does nothing for the second, use-after-free path. Serializing the interface to a single receive queue with RPS and RFS disabled removes the concurrency the race needs, but on hardware that actually spreads traffic across queues, all receive work collapses onto one CPU and network throughput under load can suffer. Vitko measured both stopgaps on his own lab VMs and found normal-traffic throughput roughly unchanged there, while cautioning that his Hyper-V setup already delivered packets to a single queue, so the single-queue cost needs measuring on real multi-queue hardware, not assumed.

A mitigation you have not measured is a hope with a sysctl in it.

What this episode teaches

Three durable points. First, kernel race conditions do not need exotic triggers: this one fires against a plain listen() on a default install, and the precondition that decides exposure, receive queues and RPS, is infrastructure configuration that most vulnerability workflows never query. Second, embargoed coordination worked as designed here: report, upstream fix, distribution embargo, then public advisory with honest severity language. The system fails often enough that it is worth marking when it does not. Third, the gap between disclosure and fleet protection is where incidents actually live. An eleven-year-old bug became public knowledge in a weekend; whether it becomes a crashed production host depends entirely on who reads a changelog, counts their receive queues, and schedules a reboot before someone else starts knocking.

Sources: Hyunwoo Kim, NoisyTCP advisory on oss-security and the V4bel/NoisyTCP repository (Nov 15-16, 2026); Red Hat CVE-2026-99288 advisory (fetched Nov 16, 2026); BleepingComputer (Nov 16, 2026); iTnews (Nov 16, 2026); Brian Vitko, getrff.com analysis (Nov 15, 2026).

📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.