Skip to content
Noisy TCP

Noisy TCP

github.com • November 15, 2026

This report is a security advisory for the NoisyTCP (CVE-2026-99288) vulnerability reported and patched by Hyunwoo Kim (@v4bel) . It is a Linux Kernel TCP Pre-auth Remote Corruption vulnerability.

To trigger this vulnerability, a stale ehash lookup and the freeing and reallocation of an existing request socket must proceed concurrently in different RX contexts. Packet processing has to be spread out by using the difference in receive hash between IPv4 fragments and regular TCP packets, so it can be triggered against servers that have two or more RX queues, or where RPS distributes packets across multiple CPUs.

That is, since this vulnerability can be triggered against "plain" TCP listeners, it can be triggered pre-auth and remotely against servers that TCP packets sent by an attacker can reach directly, such as Nginx, Apache, and SSH servers running with their "default" configuration.

The embargo agreed upon after reporting this vulnerability to [email protected] has ended, so this NoisyTCP document is now being published.

There are two trigger flows for this vulnerability, and an attacker can arbitrarily choose either one to trigger it.

1. request_sock Reuse ABA Race

While a stale lookup has found an existing request socket, another RX context frees it and reallocates a SYN cookie request with the same tuple at the same slab address.

The stale lookup acquires a refcount on the new object and then even passes the tuple recheck, but the new request has no listener attached to it. After that, the TCP receive path dereferences the NULL listener, which causes a kernel panic.

In the end, a remote DoS can be triggered.

Because this is a Linux kernel TCP vulnerability, if the DoS is triggered against a web server running in a container, it takes down not only the web server application but the entire host.

2. Full child socket ABA Race

This path can trigger a remote use-after-free, but since complex race scenarios have to succeed in a nested way, it is generally hard to trigger without using special techniques.

To trigger this vulnerability, another RX context has to get in between the first tuple check and the refcount acquisition. Because of this, against bare-metal servers, especially quiet servers with little execution variance and contention, triggering the race can take a long time.

Against Nginx, Apache, and SSH servers running with their default configuration on VMs such as QEMU/KVM, where vCPU preemption can occur, it took 5 seconds to 2 minutes to trigger.

Against Nginx running with its default configuration on a quiet bare-metal server (8 RX queues), it took 14 hours to trigger, even though a PoC with race timing carefully tuned for the target environment was used.

The affected range of NoisyTCP (CVE-2026-99288) is as follows.

Introduced: 079096f103fa (2015-10-03)

Fixed: 4e2b3ef19e0e (2026-11-03)

That is, this vulnerability had been latent in the Linux kernel for "11 years" and affects almost all Linux distributions.

Applying the 4e2b3ef19e0e patch is the priority, but if it is difficult to apply it immediately, one of the following temporary mitigations can be used.

1. Disable SYN cookies

This blocks the path where a request socket is reconstructed through a SYN cookie.

This setting is managed per network namespace. It has to be applied to "each" container that uses a separate network namespace. Disabling it lowers the protection against SYN floods.

2. Serialize to a single RX queue

This vulnerability requires different RX contexts to race concurrently. Reducing the RX queues of the interface where external packets come in to one and disabling RPS and RFS can block the race.

Extracted Entities

Domains (1)

Email Addresses (1)

Tools (2)

Vulnerabilities (1)