Security Digest · Nov 15, 2026 · Vulnerability · OS / kernel
NoisyTCP CVE-2026-99288: Remote Linux TCP Kernel Panic
Pre-auth race in Linux TCP lookups lets remote packets panic hosts since 4.4. Fixed in 6.12.117, 6.18.63, 7.3.4 and Debian. Patch this week.
Our verdict
Patch this week
Anyone who can reach a TCP port can, with patience, crash an unpatched Linux host, and virtual machines fall in minutes in the researcher's tests, but it is a crash rather than a takeover and no public exploit is known, so roll fixed kernels into this week's patch cycle.
Who must act: Teams running internet- or tenant-facing Linux servers, especially multi-queue cloud VMs and container hosts, on kernels 4.4 and later without the November 2026 fix.
Who can relax: Hosts already booted into 6.12.117, 6.18.63, 7.3.4 or a distribution kernel listing the fix, and machines that accept no inbound TCP from untrusted networks.
- Affected
- Linux kernel 4.4 and later (since commit 079096f103fa, October 2015) without commit 4e2b3ef19e0e
Linux stable 6.12 before 6.12.117, 6.18 before 6.18.63, 7.3 before 7.3.4
Longterm 5.10, 5.15, 6.1 and 6.6: no upstream fix listed in the CVE record as of Nov 16, 2026
Debian bookworm before 6.1.192-1 and trixie before 6.12.117-1
Ubuntu 16.04 through 26.10 (status 'needs evaluation' as of Nov 16, 2026)
Red Hat Enterprise Linux (Red Hat rates Important; per-release fix status not disclosed as of Nov 16, 2026) - Fixed in
- Linux 6.12.117, 6.18.63, 7.3.4 and 7.4-rc2
Debian trixie linux 6.12.117-1 (DSA-6628-1)
Debian bookworm linux 6.1.192-1 (DLA-4906-1)
Debian sid linux 7.3.4-1 - CVSS
- 7.5 (v3.1, vendor)
- EPSS
- 0.6% (percentile 49)
- CISA KEV
- Not listed
- Exploitation
- No known exploitation
- Severity
- High
What to do
- Run uname -r on every Linux host and compare it with your distribution's fixed kernel for CVE-2026-99288.
- Install the fixed kernel and reboot into it, starting with internet-facing VMs, load balancers and shared container hosts.
- Where no fixed kernel exists yet, decide per host whether to set net.ipv4.tcp_syncookies=0 in every network namespace, accepting weaker SYN flood protection.
- On hosts that can afford it, reduce the public interface to one RX queue and disable RPS and RFS as a stopgap.
- Watch your distribution's CVE page for Ubuntu and Red Hat errata, and for backports to the 5.10, 5.15, 6.1 and 6.6 longterm kernels.
How to check if you are affected or compromised
- Search kernel logs from previous boots for the oops signature: journalctl -k -b -1 | grep -E 'NULL pointer dereference|tcp_v4_rcv'.
- Treat unexplained reboots or kdump captures on internet-facing hosts since mid-November as candidate events, and check the crash RIP for tcp_v4_rcv.
- Check whether a host meets the race preconditions: ethtool -l eth0 for the RX queue count and cat /sys/class/net/eth0/queues/rx-*/rps_cpus for RPS.
- Look for sustained floods of IPv4 fragments mixed with SYNs to one port in flow logs or packet captures, which matches the published trigger method.
What happened#
On November 15, 2026 (UTC), researcher Hyunwoo Kim published NoisyTCP on oss-security after an embargo with the linux-distros list expired. CVE-2026-99288 is a race in the kernel's lockless lookup of established TCP sockets. Those sockets live in memory that can be recycled while a lookup still holds the old address. According to the CVE record, a lookup can grab a reference to a brand-new socket with the same connection tuple before it has been hashed. In Kim's proof of concept, that new object is a SYN cookie request socket with no listener, and tcp_v4_rcv() dereferences the NULL listener and panics the kernel.
The bug dates to commit 079096f103fa from October 2015, first shipped in Linux 4.4. The upstream fix, written by Kim and reviewed by Eric Dumazet and Kuniyuki Iwashima, was committed on November 3. It makes the IPv4 and IPv6 lookups restart when they land on an unhashed socket.
Who is affected#
Almost any Linux host that accepts TCP connections from an attacker. Kim's advisory says a plain listener is enough, including default nginx, Apache and SSH setups. One more condition matters: the attacker has to get packets for the same connection processed on different CPUs at once. That requires two or more NIC receive queues, or RPS spreading packets across CPUs. Kim's method mixes IPv4 fragments and regular TCP packets, which hash to different queues. Most cloud VMs and servers have multiple queues.
The CVE record lists fixes in stable 6.12.117, 6.18.63 and 7.3.4, plus 7.4-rc2. It lists no fix for the 5.10, 5.15, 6.1 or 6.6 longterm branches as of November 16. Debian has shipped fixed kernels anyway: trixie 6.12.117-1 in DSA-6628-1 and bookworm 6.1.192-1 in DLA-4906-1. The Debian tracker shows sid fixed in 7.3.4-1. Ubuntu rates the issue Medium, and every supported release from 16.04 to 26.10 still reads "needs evaluation." Red Hat rates it Important, but per-release RHEL errata were not visible to us as of November 16.
Do you really need to act?#
Yes, but on a normal patch cycle, not overnight. No authentication is needed, and on a container host one crash takes down every tenant. That is enough to justify patching this week.
Three facts keep this below emergency level. First, the confirmed impact is a crash. Kim describes a second path that could lead to a remote use-after-free, but he calls it harder to trigger and says there is no general-purpose exploit and "not yet" remote code execution. Red Hat says it allows no data change or code execution. Second, timing varies widely. Kim triggered the crash in 5 seconds to 2 minutes on QEMU/KVM guests, but a quiet bare-metal server with 8 receive queues running nginx took about 14 hours. Third, no in-the-wild exploitation or public exploit has been reported. Kim mentions only a limited, lab-specific exploit.
Red Hat's preliminary CVSS 3.1 score is 7.5. Its vector rates confidentiality and integrity High, which does not match its own statement that the flaw is denial of service only. NVD had not scored the CVE as of November 16.
What to do now#
- Inventory kernels with
uname -rand compare against your distribution's CVE page. - Install the fixed kernel and reboot, starting with internet-facing VMs, load balancers and multi-tenant container hosts. On Debian:
apt update && apt install linux-image-amd64, then reboot. - Where no fixed kernel exists yet, the researcher and Red Hat offer two stopgaps. You can disable SYN cookies with
sysctl -w net.ipv4.tcp_syncookies=0. That setting is per network namespace, so apply it inside each container with its own namespace, and it weakens SYN flood defense. Or you can drop the public interface to one receive queue (ethtool -L eth0 combined 1) and disable RPS and RFS, which costs throughput. - Track Ubuntu, Red Hat and longterm-kernel backports, and remove the stopgaps once you have booted a fixed kernel.
How to tell if you were hit#
A successful attack is a kernel panic, so look for unexplained reboots. Check previous boots with journalctl -k -b -1 | grep -E 'NULL pointer dereference|tcp_v4_rcv' and inspect any kdump captures for a fault in tcp_v4_rcv. To see whether a host meets the race conditions, run ethtool -l eth0 and cat /sys/class/net/eth0/queues/rx-*/rps_cpus. Flow logs showing sustained IPv4 fragments mixed with SYNs to one port fit the published method. The bug crashes the host and leaves no persistent foothold, so recovery is a reboot into a fixed kernel.
Timeline#
- — Commit 079096f103fa introduces the bug, first shipped in Linux 4.4.
- — Fix 4e2b3ef19e0e committed upstream.
- — CVE-2026-99288 published by the kernel CNA.
- — Debian ships DSA-6628-1 for trixie.
- — Debian LTS ships DLA-4906-1 for bookworm. Embargo ends and Kim publishes NoisyTCP on oss-security.
- — BleepingComputer reports the flaw. Ubuntu and Red Hat fixes not yet visible.
Vulnerability records
- CVE-2026-99288 · In the Linux kernel, the following vulnerability has been resolved: inet: reject unhashed sockets during established lookups TCP established ehash lookups are…
Sources
Original sources first: vendor and project advisories, CVE and government records, the discoverers’ write-ups. Every claim above links to one of these. See how the Security Desk works.
Original and primary sources
- Hyunwoo Kim via oss-security — NoisyTCP: Linux Kernel TCP Pre-auth Remote Corruption (CVE-2026-99288) openwall.com
- CVE record — CVE-2026-99288 cve.org
- Linux kernel — inet: reject unhashed sockets during established lookups git.kernel.org
- Hyunwoo Kim — NoisyTCP advisory repository github.com
- Red Hat — CVE-2026-99288 access.redhat.com
- Debian security tracker — CVE-2026-99288 security-tracker.debian.org
- Debian — DSA-6628-1 linux security update lists.debian.org
- Debian LTS — DLA-4906-1 linux security update lists.debian.org
- Ubuntu — CVE-2026-99288 ubuntu.com
Coverage and analysis
- BleepingComputer — New NoisyTCP vulnerability lets attackers crash Linux servers bleepingcomputer.com