A vulnerability in the Linux kernel's TCP handling allows an attacker to crash a machine over the network. Security researcher Hyunwoo Kim, who found it, has named it "NoisyTCP," and it is tracked as CVE-2026-99288. Machines running Nginx, Apache or an SSH server with default settings can be targeted without any authentication. With the coordinated embargo among Linux distributors now over, Kim has published a technical write-up on GitHub.
Crafted TCP Packets Can Trigger a Kernel Panic
According to Kim's NoisyTCP write-up and Red Hat's vulnerability page, the bug is a race condition in the code that looks up TCP connections (lockless socket lookups). Recycled socket memory can be accessed before it has been properly initialized and hashed, which leads to a NULL pointer dereference and a kernel crash.
An attacker only needs to send crafted TCP packets; no login or other authentication is required. Red Hat rates the issue "Important," the second-highest of its four severity levels, with a preliminary CVSS v3 base score of 7.5. At the same time, Red Hat says the flaw does not allow data alteration or arbitrary code execution. Asked whether it could lead to code execution, Kim likewise answers that it cannot at present and basically should not.
Servers That Spread Receive Processing Across CPUs Are Affected
The race requires a server with two or more network receive queues (RX queues), or with RPS (a mechanism that distributes incoming packets across multiple CPUs) enabled. Kim says the attack exploits the difference in receive hash between IPv4 fragments and regular TCP packets to get them processed on different CPUs.
How long an attack takes varies widely. In Kim's testing, QEMU/KVM virtual machines crashed within about 5 seconds to 2 minutes. By contrast, a quiet bare-metal server with eight receive queues took about 14 hours.
Introduced in 2015, Fixed on November 3, 2026
The write-up traces the bug to a commit from October 3, 2015, and the fix to a commit dated November 3, 2026. The fix, titled "inet: reject unhashed sockets during established lookups," makes the kernel restart the lookup when it hits a socket that has not been hashed yet. Kim says the bug affects almost all Linux distributions.
Distributions are rolling out updates. The Debian security tracker lists the issue as fixed in 6.1.192-1 for Debian 12 (bookworm) and 6.12.117-1 for Debian 13 (trixie). The development branch sid is fixed in 7.3.4-1, while testing (forky) still carried an unfixed version as of November 16.
Workarounds If You Cannot Update Right Away
The real fix is to update to a kernel that includes the patch and reboot. For systems that cannot be updated immediately, Kim and Red Hat list two workarounds.
- Disable SYN cookies (
sysctl -w net.ipv4.tcp_syncookies=0). The setting applies per network namespace, and it weakens protection against SYN flood attacks - Reduce the receiving interface to a single RX queue and disable RPS and RFS so the race cannot occur
Both trade away performance or defenses, so they should be treated only as stopgaps until the update is applied.
Our Take
NoisyTCP does not let anyone take over a server, but the fact that anyone can remotely knock out a machine exposing a web server or sshd is not something to dismiss. Given that it reproduced within minutes on virtual machines, we think services running in the cloud in particular should update soon. With a detailed write-up already public, administrators should also expect attackers to try it.
Summary
NoisyTCP (CVE-2026-99288) exploits a race condition in the Linux kernel's TCP code to trigger a kernel panic without authentication. It cannot alter data or execute code, but servers with multiple receive queues are broadly exposed. The fix was merged on November 3, 2026, and installing your distribution's updated kernel and rebooting is the reliable remedy. If you cannot update right away, disabling SYN cookies or reducing to a single receive queue can serve as a temporary measure.