CVE-2026-99288: NoisyTCP fixes for Linux servers
CVE-2026-99288, dubbed NoisyTCP, is a Linux kernel race that lets an unauthenticated attacker crash a host through a reachable TCP listener. Debian lists fixes in 6.1.192-2, 6.12.117-1 and 7.3.4-1; Ubuntu fixes remain pending. Ubuntu 20.04’s linux-hwe-5.8 package won’t be fixed, and disabling SYN cookies as a workaround requires applying it in each separate network namespace.
Summary written with AI from this article, and checked against it.

At a glanceCVE-2026-99288
CVE-2026-99288 has no CVSS score yet. No fixed version has been published yet. Debian bookworm fixed it in linux 6.1.192-2.
| Severity | Not scored yet (NVD: Received) |
|---|---|
| Fixed in | No fixed version published yet |
| Affected | Not published yet |
| Exploited | Not in CISA's Known Exploited Vulnerabilities catalog |
| EPSS | 0.2% chance of exploitation in the next 30 days (FIRST, November 16, 2026) |
| Public exploit | 3 GitHub repositories name it: proofs of concept, scanners, detections. Unverified: some are fake or malicious. |
| Ubuntu | 26.04 resolute linux: Vulnerable, fix pending · 24.04 noble linux: Vulnerable, fix pending · 22.04 jammy linux: Vulnerable, fix pending · 20.04 focal linux: Vulnerable, fix pending · 20.04 focal linux-hwe-5.8: Won't be fixed · 18.04 bionic linux: Vulnerable, fix pending |
| Debian | bookworm linux: Fixed (6.1.192-2) · trixie linux: Fixed (6.12.117-1) · forky, sid linux: Fixed (7.3.4-1) |
| Red Hat | RHEL 10: Affected, RHEL 8: Affected, RHEL 8: Affected |
Is there a public exploit for CVE-2026-99288?
Public code exists as of November 17, 2026: 3 public GitHub repositories name it (proofs of concept, scanners and detections). It is not in CISA's Known Exploited Vulnerabilities catalog. None of it is verified: some such repositories are fake or carry malware, so read the code before running it, and only against systems you own. FIRST's EPSS puts the chance of exploitation in the next 30 days at 0.2%, higher than 4.8% of scored CVEs.
From NVD, Ubuntu, Debian, Red Hat · checked November 17, 2026
An attacker can crash an entire Linux host through a TCP listener, even when the application runs in a container. The flaw is CVE-2026-99288, dubbed NoisyTCP, and the race is in the kernel. Debian lists bookworm as fixed in 6.1.192-2; Ubuntu still lists fixes as pending. The first job is to compare your kernel package with your distribution’s tracker, then decide whether you can patch or need a temporary mitigation.
BleepingComputer reports that researcher Hyunwoo Kim discovered CVE-2026-99288, which is reachable through default Nginx, Apache and SSH configurations without authentication. This is not just a failed request or a service that needs restarting. The kernel can panic and take the host down with it. The reported impact is loss of availability, not data alteration or arbitrary code execution.
Check CVE-2026-99288 exposure on Linux servers
Start with the kernel and the path packets take to it, rather than the application’s version number. Kim describes the vulnerable target as a Linux server that directly receives the attacker’s TCP packets. The race also needs concurrent receive processing. That can come from multiple receive queues or from RPS distributing packets across CPUs.
Kim reports crashes against default Nginx, Apache and SSH configurations on virtual machines such as QEMU/KVM in about 5 seconds to 2 minutes. His carefully tuned proof of concept took about 14 hours against default Nginx on a quiet bare-metal server with 8 receive queues.
I’d put directly reachable production listeners first, especially those on virtual machines. Sorting the work by whether an application uses PHP, Python or Node.js would miss the relevant distinction: CVE-2026-99288 sits below all of those application choices.
Compare the running kernel’s package with your distribution’s security status. The Debian package entries below give you that comparison where applicable. On Ubuntu, keep a pending fix separate from an installed one. Finding a CVE entry is not, by itself, evidence that your host has received a corrected kernel.
If you cannot apply an update immediately, assess one of the documented temporary mitigations. Red Hat’s advisory gives the following command for disabling SYN cookies. This is a networking change, not a version check, and it comes with a cost: reduced protection against SYN floods.
sudo sysctl -w net.ipv4.tcp_syncookies=0The setting applies per network namespace. Red Hat says you must also apply it in each container that uses a separate network namespace; setting it on the host alone does not cover those containers.
The other documented mitigation reduces the externally facing interface to a single receive queue and disables RPS and RFS. Neither advisory supplies the interface-specific commands needed to make that change, so there is no command sequence to copy here. The receive-processing conditions matter, as do the trade-offs described below.

Which Linux kernel packages contain the NoisyTCP fix?
Debian’s security tracker identifies fixed packages for bookworm, trixie, forky and sid. Read the package rows, not just the release names: the current listing separates vulnerable release packages from corrected packages in the security channel.
| Release | Version | Status |
|---|---|---|
| bookworm | 6.1.176-1 | vulnerable |
| bookworm (security) | 6.1.192-2 | fixed |
| trixie | 6.12.107-1 | vulnerable |
| trixie (security) | 6.12.117-1 | fixed |
| forky, sid | 7.3.4-1 | fixed |
Debian records 6.1.192-1 as bookworm’s first-fixed version, while its current security-package listing and the distribution status checked on November 17, 2026 identify 6.1.192-2 as fixed. Use the current security-channel package as the operational target rather than seeking out the earlier revision.
Ubuntu’s tracker still marks the main linux package as vulnerable, with fixes pending, for 26.04, 24.04, 22.04, 20.04 and 18.04. Pending means distribution work has not yet been completed in those entries; it does not mean an update has already reached the host. Debian’s version numbers cannot stand in for that check, either. Each tracker identifies packages for its own distribution.
There is also a difference between an upstream correction existing and a distribution package containing it. Kim identifies the upstream patch for CVE-2026-99288 as commit 4e2b3ef19e0e. That gives you an upstream correction to refer to, but its existence alone says nothing about whether every distribution package has incorporated the change.
Why a default TCP listener can bring down the host
The CVE-2026-99288 race lets a stale socket lookup accept a replacement object that is not ready to be returned. In Kim’s technical advisory, an established-socket lookup races with another receive context that frees and reallocates a request socket. The lookup can end up working with the replacement rather than the object it originally found.
Checking the connection tuple again does not necessarily expose that substitution. In the documented trigger, the replacement occupies the same memory address and carries the same tuple. What it does not have is the listener that the receive path expects to find next.
- A lookup keeps the address of an existing request socket while processing continues in another receive context.
- That other context frees the socket, then reuses its memory for a SYN-cookie request socket with the same tuple.
- The stale lookup acquires a reference to the replacement, which passes the tuple recheck. The TCP receive path then follows a NULL listener pointer and causes the kernel panic.
Debian’s technical description spells out the correction. After acquiring a reference, established lookups check whether the socket is unhashed. If the object is unhashed, or its tuple no longer matches, the lookup releases the reference and starts again. Related request-socket failure paths also release references instead of directly freeing an object that another lookup might still reference.
Application authentication is therefore not the boundary that stops this failure. Kim says ordinary TCP listeners are enough to reach it. Nginx, Apache and SSH are examples of reachable services here, not three separate application vulnerabilities that need three separate application patches.
Those timings describe the researcher’s tests, not an attack duration you can count on. Kim points to virtual-CPU preemption as one contributor to the virtual-machine results. Quiet bare metal can make the race harder to win, but the successful bare-metal test is still a successful test. It does not support marking that configuration as unaffected.
NoisyTCP workarounds leave networking trade-offs
-
Disabling SYN cookies gives up one protection to temporarily block the documented trigger path. Kim and Red Hat both explain the mechanism: it prevents a request socket from being reconstructed through a SYN cookie. The same change lowers protection against SYN floods, so that cost belongs in the decision rather than in a footnote.
The network-namespace scope is easy to underestimate. Under the advisory’s instructions, a host-level setting does not cover every container with a separate network namespace. Before relying on this CVE-2026-99288 workaround, account for the namespaces that receive the relevant traffic and check where the setting has actually been applied.
-
The receive-queue alternative removes a different prerequisite: concurrency between receive contexts. Red Hat specifies one receive queue on the incoming interface, with RPS and RFS disabled. Kim’s advisory presents that change and disabling SYN cookies as alternative temporary mitigations. It does not say that you must apply both alternatives together.
Reducing the queue count is not completion of the documented procedure while RPS or RFS remains enabled. The interface matters, too. Changing an interface that does not receive the external packets would leave the change outside the scope specified by the advisory.
I’d take a distribution-provided patched kernel over either networking workaround as a permanent answer. Until a patch is available, the choice is more specific: assess the SYN-flood protection you would give up, or assess whether the complete receive-serialization change is appropriate for that host. These are kernel-risk mitigations with networking consequences. Neither one is an application fix.
Prioritise host availability, not exploit popularity
The evidence supports treating CVE-2026-99288 as a remotely reachable denial-of-service risk. It does not support adding a claim that an attack campaign has been confirmed. Red Hat rates the impact as Important and describes an unauthenticated attacker causing an operating-system crash across the network.
As of November 17, 2026, NVD’s CVE record has neither a published CVSS score nor a fixed version published there. That gap is not a reason to ignore the fixed distribution packages Debian identifies. Use distribution trackers, rather than NVD’s publication status, to decide which packages contain fixes. Those are different checks.
CISA does not list CVE-2026-99288 in its known-exploited catalog in the checked status. That absence does not prove nobody has attacked it. PoC-in-GitHub, separately, identifies public repositories naming the flaw but labels those listings as unverified. A repository appearing in that list does not establish that it contains a reliable, general-purpose exploit.
Kim describes his own exploit as limited and lab-specific, targeting virtual-machine servers, and says general-purpose exploit research is still in progress. Keep that qualification next to the successful crash results. It is part of the evidence, not a reason to discard the crashes or to describe the exploit as more general than he does.
I’d put exposed virtual-machine hosts near the front of the patch queue, then work through the other directly reachable listeners and their receive-processing configuration. Container boundaries and application login screens do not address this kernel failure. For each host, record a fixed package or a temporary mitigation whose scope matches the traffic and network namespaces involved.
Frequently asked questions
When was NoisyTCP introduced into the Linux kernel?
Hyunwoo Kim identifies commit 079096f103fa, dated October 3, 2015, as the introduction point. His advisory dates the correcting commit to November 3, 2026.
Will Ubuntu 20.04’s linux-hwe-5.8 package receive a NoisyTCP fix?
Ubuntu’s tracker marks the Ubuntu 20.04 linux-hwe-5.8 entry as “Won’t be fixed.” That is a separate package entry from Ubuntu 20.04’s main linux package, which is marked vulnerable with a fix pending.
Sources
Primary sources
- Distribution trackerCVE-2026-99288 · access.redhat.com
- CodeGitHub - V4bel/NoisyTCP · github.com
- Distribution trackerCVE-2026-99288 · security-tracker.debian.org
- CVE recordNVD and distribution data for CVE-2026-99288 (checked 2026-11-17) · nvd.nist.gov
Reports of the story
- News reportNew NoisyTCP vulnerability lets attackers crash Linux servers · bleepingcomputer.com
How this article was made: drafted with AI from the sources above and fact-checked against them, then reviewed by Vishal on . How we write and check articles.
