11-Year-Old NoisyTCP Linux Kernel Bug Can Crash Cloud VMs in 5 Seconds, With RHEL and Ubuntu Patches Pending

11-Year-Old NoisyTCP Linux Kernel Bug Can Crash Cloud VMs in 5 Seconds, With RHEL and Ubuntu Patches Pending

An 11-year-old vulnerability in the Linux kernel’s TCP networking stack, tracked as CVE-2026-99288 and dubbed NoisyTCP, allows unauthenticated remote attackers to crash internet-facing servers using crafted network traffic.

While upstream Linux maintainers have already released a fix, official kernel updates for Red Hat Enterprise Linux and Canonical Ubuntu remain pending as of November 17, 2026, leaving large numbers of cloud and enterprise systems reliant on temporary workarounds.

Security researcher Hyunwoo Kim published the full technical advisory on the Openwall oss-security mailing list at 16:00 UTC on November 15, 2026, following the expiration of a coordinated disclosure embargo with Linux distribution maintainers. According to Kim’s advisory, the flaw was introduced on October 3, 2015, in kernel commit 079096f103fa and first shipped in Linux kernel 4.4 in January 2016. That timeline puts more than a decade of Linux server releases within the affected window.

Unlike network vulnerabilities that require unusual protocol extensions or login credentials, NoisyTCP targets standard TCP listeners. Any Linux server that accepts incoming TCP connections and distributes packet processing across two or more receive (RX) queues or multiple processor cores via Receive Packet Steering (RPS) is potentially exposed. Kim noted in his advisory that the flaw can be triggered remotely without authentication against servers that attacker packets can reach directly, including “Nginx, Apache, and SSH servers running with their ‘default’ configuration”.

How quickly an attacker can trigger the crash depends heavily on where the target system runs. In laboratory testing, Kim found that default Nginx, Apache, and SSH setups running inside QEMU/KVM virtual machines suffered a kernel panic in roughly five seconds to two minutes.

By contrast, a quiet bare-metal server with eight RX queues held out for approximately 14 hours, even when targeted with a proof-of-concept exploit whose timing was tuned specifically for that hardware.

Those figures reflect specific research benchmarks rather than a guaranteed attack timeline across every cloud provider. However, they highlight why virtualized infrastructure is uniquely sensitive to race conditions. Triggering NoisyTCP requires one packet-processing thread to interrupt another inside a timing window measured in nanoseconds. On cloud virtual machines, hypervisor vCPU preemption can pause a virtual processor mid-operation, widening that window and making the race far easier to win.

Containerized workloads also fail to isolate the blast radius. Because containers share the underlying host operating system’s kernel networking stack, crashing a single web server running inside a Docker container or Kubernetes pod brings down the entire host machine and every other workload running on it.

Technically, NoisyTCP resides in the code Linux uses to perform lockless lookups in its established TCP connection hash table (ehash), which relies on SLAB_TYPESAFE_BY_RCU memory caches to recycle socket structures quickly. By mixing fragmented IPv4 packets with regular TCP segments to exploit differences in receive hashing, an attacker can force incoming packets onto separate receive queues running concurrently.

In the primary attack path, one receive thread looks up an existing connection request socket while a second thread frees that memory and reallocates a TCP SYN cookie request socket with the same network tuple at the exact same memory address. SYN cookies are a standard defense mechanism that lets Linux handle connection floods without exhausting memory.

According to the upstream Linux kernel commit description, the first thread acquires a reference to the newly allocated SYN cookie socket and passes its tuple recheck without verifying whether the socket is actually hashed. Because SYN cookie request sockets carry a NULL listener pointer (rsk_listener), the kernel subsequently reads through a NULL pointer inside tcp_v4_rcv() and triggers an immediate kernel panic.

Kim also documented a second, more complex trigger path involving a nested race condition on full child sockets that causes a remote use-after-free memory error. However, that scenario was observed on a modified test kernel with artificial delays, was questioned by the upstream Linux TCP maintainer as to whether it can occur on an unmodified production kernel, and was omitted from the final upstream commit description. Kim himself stated that general-purpose remote code execution has not been achieved beyond a limited lab exploit on virtual machines, no independent reproduction of the second path has been published, and Red Hat Product Security confirmed that the flaw does not allow data alteration or arbitrary code execution.

Red Hat’s public advisory does contain a notable scoring discrepancy that security teams should account for when prioritizing remediation. Red Hat rates CVE-2026-99288 as Important with a CVSS v3.1 score of 7.5, using a vector string (CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:H/I:H/A:H) that indicates low privilege requirements alongside high confidentiality and integrity impact. That vector conflicts with both Red Hat’s own written impact statement and the upstream kernel commit description, which classify the issue as an unauthenticated denial-of-service crash with no data modification or code execution.

Upstream Linux maintainers merged Kim’s primary fix, commit 4e2b3ef19e0e (“inet: reject unhashed sockets during established lookups”), on November 3, 2026, and shipped it in stable kernel releases 6.12.117, 6.18.63, and 7.3.4 on November 10. Organizations running ARM-based cloud instances, such as AWS Graviton or Ampere servers, as well as POWER and RISC-V systems, face an additional architecture-specific detail.

On CPUs with weak memory ordering, a separate upstream commit, 5b5d8f351dd9 (“net: restore refcount ordering for lockless socket access”), is needed alongside 4e2b3ef19e0e to prevent socket state reads from being reordered before reference-count acquisition, and that follow-up fix was queued for later stable kernel updates such as 7.3.5 and 7.4-rc2 rather than the initial November 10 releases.

Distribution patch availability remains divided as of November 17, 2026. Debian has published updated kernels for Debian 12 Bookworm (6.1.192-1 and 6.1.192-2 under DLA-4906-1) and Debian 13 Trixie (6.12.117-1 under DSA-6628-1), though Debian’s forky testing branch remains unpatched. Oracle Linux released patched Unbreakable Enterprise Kernel updates (ELSA-2026-500630, 500631, and 500632) on November 12 and 13, Arch Linux has updated to patched 6.18.63 and 7.3.4 kernels, and CIQ has issued updates for its commercial Rocky Linux builds.

By contrast, Red Hat lists Red Hat Enterprise Linux 8, 9, and 10, along with OpenShift Container Platform 4 and 5, as affected with no errata published yet, while older RHEL 6, RHEL 7, and CentOS 7 releases predate the 2015 code change and are unaffected. Canonical’s Ubuntu security tracker lists all supported Ubuntu releases from 16.04 through 26.10, including 22.04 LTS and 24.04 LTS, as still under evaluation without patched packages.

Community builds of AlmaLinux and Rocky Linux also lacked fixes as of November 17, and TuxCare confirmed on November 16 that rebootless KernelCare live patches are not yet available for any distribution, meaning administrators should prepare for a standard kernel package update and host reboot once their vendor ships a patch.

Administrators should not rely solely on upstream kernel version numbers to judge exposure, because distributions backport security patches into older kernel series. Debian 12’s 6.1.192-1 package includes the NoisyTCP fix even though upstream Linux 6.1.192 does not, making distribution security trackers and package changelogs the only reliable way to confirm patch status. Operators who applied last month’s Linux TCP patch for CVE-2026-98432 should also note that NoisyTCP is a separate race condition requiring its own update.

No confirmed in-the-wild exploitation of CVE-2026-99288 has been reported as of November 17, 2026, and the flaw is not listed in CISA’s Known Exploited Vulnerabilities catalog. Even so, because a successful attack crashes the kernel abruptly without leaving an application-level warning, teams running unpatched internet-facing Linux virtual machines should check whether multi-queue networking (ethtool -l) or RPS is enabled and inspect kernel crash logs (kdump or pstore) for unexpected tcp_v4_rcv faults.

Where vendor updates are unavailable, both Kim and Red Hat outline two temporary workarounds, each carrying clear operational trade-offs. Disabling TCP SYN cookies via sudo sysctl -w net.ipv4.tcp_syncookies=0 blocks the primary NULL pointer kernel-panic path, but it leaves servers far more exposed to classic SYN flood denial-of-service attacks.

Because tcp_syncookies is managed per Linux network namespace, administrators must also apply that setting separately inside every container that uses its own network namespace. If operators choose this stopgap, pairing it with a larger net.ipv4.tcp_max_syn_backlog queue and upstream firewall or edge SYN rate-limiting can help absorb modest connection spikes, though neither measure replaces full SYN cookie protection on heavily targeted hosts.

Alternatively, reducing the external network interface’s receive queues to one and disabling both RPS and Receive Flow Steering (RFS) serializes packet processing and blocks concurrent race conditions entirely, though it can create network throughput bottlenecks on busy multi-core servers.

More from Vulnerabilities

Community discussion

Join the conversation

0 comments

Be the first to join in

There are no comments yet. Sign in to start a thoughtful discussion.