Monday 16 November 2026
Security

NoisyTCP flaw lets attackers crash Linux servers remotely

An unauthenticated attacker can crash default Nginx, Apache and SSH servers with crafted packets. The 11-year-old kernel race has patches upstream, but many distributions have not shipped them.

Photograph of the large Tux the Linux penguin figure at IBM.
A physical Tux the Linux penguin figure, the Linux mascot, photographed at IBM in Brazil. Photograph: Renato.pierri/Wikimedia Commons (CC BY-SA 4.0)

BleepingComputer reported that CVE-2026-99288, a Linux kernel race condition named NoisyTCP, lets unauthenticated attackers crash servers running default Nginx, Apache or SSH. Researcher Hyunwoo Kim found the flaw and said almost all Linux distributions carry it. Kim said crafted TCP packets exploit timing between SYN cookie handling and socket hashing. The crash causes a kernel panic and takes down the host, but it does not allow altering data or arbitrary code execution.

Cyber Kendra reported that the flaw was introduced on October 3, 2015, and CybersecurityNews reported that the introductory commit was 079096f103fa and the fix commit 4e2b3ef19e0e landed on November 3, 2026. The Linux kernel CVE team released the record on Tuesday, November 10 with no CVSS score, and NVD had not scored it as of November 15.

How the race crashes the kernel

CybersecurityNews reported on Monday that Kim's technical write-up describes a race in the kernel's TCP socket lookup. When a TCP packet arrives, the kernel searches for the matching socket in the established-connections hash table without acquiring a lock. Memory for freed sockets is marked SLAB_TYPESAFE_BY_RCU, so a freed request socket's memory can be reused while another CPU still looks at the old address. The kernel normally checks the connection tuple after taking a reference, but CVE-2026-99288 fails to also check whether the object is still hashed.

An attacker needs only a plain TCP listener. The target must process packets on more than one CPU, either through a network interface with two or more receive queues or with Receive Packet Steering enabled. Kim's proof of concept uses IPv4 fragments and regular TCP packets that hash to different receive queues. One CPU finds an existing half-open request socket. Another CPU releases that record and immediately reallocates the same memory for a fresh SYN cookie request carrying identical connection details. The first CPU passes tuple checks because the new object has the same addresses and ports, but the new request socket has no listener attached. The function tcp_v4_rcv() then reads the connection state by way of the NULL listener pointer, causing the kernel to panic inside an interrupt handler.

Timing varies. On QEMU/KVM virtual machines, default Nginx, Apache and SSH servers crashed in about 5 seconds to 2 minutes. On a quiet bare-metal server with eight receive queues, a tuned proof of concept took about 14 hours. Cloud and virtualized servers sit on the fast side of that range. Containers provide no isolation because the crash is in the host kernel; an attack against one containerized web server takes down the host and every other container on it.

Patch availability and workarounds

The upstream patches are part of Linux 6.12.117, 6.18.63 and 7.3.4, and the mainline kernel incorporated the fix two weeks before publication. Cyber Kendra found Oracle shipped fixed Unbreakable Enterprise Kernel packages, Debian had the fix only in unstable, and Ubuntu listed no fixed package for any release as of November 15. The kernel CVE record lists no fixed release for the 5.10, 5.15, 6.1 and 6.6 long-term branches.

Where no fixed package exists, Kim recommends two stopgaps. Disable SYN cookies with sudo sysctl -w net.ipv4.tcp_syncookies=0, applied per network namespace; this removes protection against SYN flood attacks. Or reduce the network interface to a single receive queue and disable RPS and Receive Flow Steering, which blocks the race at the cost of network throughput. Teams running public Linux servers should check kernel versions, test vendor patches, and watch for unexplained reboots.

No vendor has reported attacks exploiting CVE-2026-99288, and no public exploit code had been indexed as of November 15. The CVE record confirms a proof of concept exists and reproduces the crash, including a kernel oops from a 7.3.0 test kernel, but Kim's GitHub repository describes the proof of concept without releasing code. Kim does not claim remote code execution; a second trigger path could cause a remote use-after-free, but it needs several races in a row and is much harder.

The disclosure follows other Linux kernel work by Kim. In May he disclosed Dirty Frag, a local privilege escalation chain; in July he disclosed Januscape, a 16-year-old virtual-machine escape. CybersecurityNews reported that attackers have been scanning for CVE-2026-98432, a critical TCP/IPv4 flaw rated 9.8. Cyber Kendra noted that CVE-2026-99288 is the second remotely exploitable TCP flaw introduced in Linux 4.4 to come to light within a month.

Sources: BleepingComputer, CybersecurityNews, Cyber Kendra

Noor Haddad

Noor Haddad covers attacks, breach disclosures and exploited vulnerabilities for Five Nines. Noor writes from CISA and vendor advisories, SEC disclosures and published reporting.

Corrections. Report an error to corrections@thefivenines.com. We list substantive corrections on the corrections page.