Researcher Hyunwoo Kim has disclosed NoisyTCP, tracked as CVE-2026-99288, a pre-authentication flaw in the Linux kernel’s TCP stack that lets a remote attacker crash servers running default settings. The bug is 11 years old, introduced by a 2015 commit, and affects almost all Linux distributions.
How NoisyTCP Triggers a Kernel Panic in the TCP Stack
CVE-2026-99288 is a race condition in lockless socket lookups. Under certain timing, recycled socket memory can be used before it has been initialized and hashed, which causes a NULL pointer dereference and a kernel panic. Attackers exploit the window with crafted TCP packets that hit the timing gap between SYN cookie handling and socket hashing.
The flaw needs no special setup on the target beyond a listening socket. It does require two or more receive queues, or Receive Packet Steering spreading packets across CPUs. Servers running Nginx, Apache or SSH with default configurations are affected.
Denial of Service Only, With a Harder Use-After-Free Path
The practical result is a denial of service. The flaw does not change data and does not give an attacker arbitrary code execution. Kim also described a second code path that could lead to a use-after-free condition, but that path is hard to trigger.
Exploit Timing From Seconds to Fourteen Hours
Exploitation time varies widely with the environment. Against QEMU/KVM virtual machines, the crash takes between about five seconds and two minutes. Against a quiet bare-metal Nginx server, it takes about 14 hours. A crash in a containerized web server brings down the entire host, because the fault sits in the shared kernel rather than in the container.
Patched Kernel Versions and Current Exploitation Status
The fix was merged into the mainline kernel around November 2 and is included in kernels 6.12.117, 6.18.63, 7.3.4 and 7.4-rc2. Vendors including Red Hat are tracking backports for their distributions.
Kim has published a write-up but no public exploit code. No attacks in the wild have been reported, and the flaw is not in the Cybersecurity and Infrastructure Security Agency’s Known Exploited Vulnerabilities catalog.
NoisyTCP Workarounds and Their Security Trade-Off
For systems that cannot be patched immediately, two mitigations are available. Setting net.ipv4.tcp_syncookies to 0 closes the race but weakens protection against SYN flood attacks. Alternatively, limiting external interfaces to a single receive queue and disabling RPS and RFS removes the multi-CPU condition the exploit depends on.
Hyunwoo Kim’s Disclosure Sequence for CVE-2026-99288
The mainline fix landed about two weeks before the public advisory, which gave distribution maintainers lead time to prepare backports before technical details circulated. Kim’s published write-up describes the mechanism in detail but includes no working exploit code, so an attacker would have to build tooling from the description. The affected configuration is also specific: the attack depends on a listening socket and on packets being spread across CPUs, so a server limited to a single receive queue without RPS does not meet the condition.
Kim’s Third Linux Kernel Disclosure This Year
The same researcher previously disclosed Dirty Frag in May and Januscape in July. NoisyTCP is the third kernel-level finding credited to Kim in 2026, a cadence that places sustained scrutiny on code paths that have gone largely unexamined for years.
Why an 11-Year-Old Race Condition Matters for Cloud Hosts
The analysis point is scale rather than severity. A denial of service rated lower than code execution can still carry weight when the target is the kernel on internet-facing hosts, because a single remote actor can take down a machine without credentials. The container case sharpens that: operators who treat container boundaries as isolation will find that a crafted packet stream aimed at one web service removes every workload on the node.
The gap between exploit times is itself notable: about 14 hours on quiet bare metal, but seconds on virtual machines, which leaves little room for reaction on the virtualized hosts that make up much of modern hosting. Because the fix has been in mainline for roughly two weeks, the disclosure shifts the question to how quickly distribution vendors ship the backports that most production servers actually run.
