Decoding the world of cybersecurity

NoisyTCP flaw exposes Linux servers to remote crashes

A newly disclosed Linux TCP vulnerability can allow unauthenticated remote traffic to crash affected servers. The NoisyTCP flaw involves a kernel race condition and requires distribution-specific patch checks.

NoisyTCP flaw exposes Linux servers to remote crashes
Summary
  • NoisyTCP, tracked as CVE-2026-99288, affects the Linux kernel’s handling of incoming TCP connections.
  • The researcher demonstrated a remotely triggered kernel crash without authenticating to the target service, although practical timing depends on the environment.
  • A kernel fix was committed on 3 November; operators must check whether their deployed distribution has incorporated the correction.

Security researcher Hyunwoo Kim has disclosed a vulnerability in the Linux kernel that can cause a remotely reachable server to crash before an attacker has authenticated to the service. The flaw, named NoisyTCP and tracked as CVE-2026-99288, concerns the way the kernel processes incoming TCP connections when concurrent network activity reaches the same connection state. Kim published the advisory after an agreed disclosure embargo expired.

The affected code sits below an individual application, in the operating system’s network stack. Web servers, SSH services and other ordinary TCP listeners can therefore be exposed when they run on vulnerable kernels and can receive traffic from an external source. The advisory does not suggest that the application itself must contain an exploitable bug or that an attacker must first obtain a user account. Exposure depends instead on the kernel build, network configuration and the conditions needed to trigger the concurrency error.

Kim describes a race involving an earlier connection lookup and the subsequent release and reuse of a kernel object. If processing proceeds against a state that is no longer valid, the kernel can attempt to access a missing internal reference and panic. That produces a denial-of-service condition: the affected host may stop running until it is restarted or otherwise recovered. The underlying memory-safety problem is technically more extensive, but the public demonstration establishes a crash, not reliable remote execution of attacker-controlled code.

The defect depends on events occurring at nearly the same time across network-processing contexts. In the researcher’s virtualised testing, the condition was reached considerably faster than in a quiet physical-server environment. These observations are relevant to the practical risk, although they should not be turned into a universal estimate of how quickly an arbitrary production host could be disrupted. Traffic patterns, scheduling, kernel version and hardware arrangement can all change the likelihood of hitting the race.

Container deployments can experience wider outages because their workloads normally share the host operating-system kernel. Containers running on one Linux host generally share its kernel, even when their application processes and filesystems are isolated. A kernel panic triggered while traffic reaches one containerised service can therefore affect workloads outside that container. An operator who patches only the web server package or rebuilds an application container has not necessarily removed the exposure in the shared host kernel.

Kim identifies a relevant Linux code change introduced in October 2015 and a corrective kernel commit dated 3 November 2026. That upstream history does not, on its own, establish whether a particular installed enterprise distribution is vulnerable or fixed. Maintainers frequently backport targeted corrections into older supported kernel branches, leaving their reported version numbers different from those of the mainline project. Distribution security notices and installed package histories are the appropriate sources for determining patch status.

The same vulnerable TCP processing path can be present beneath otherwise unrelated application stacks across a large Linux estate. A vulnerability in a common network path can affect infrastructure that otherwise has very different application stacks: websites, internal administration systems, virtual machines and platforms supporting customer services. Asset inventories that classify risk by application name alone may overlook a shared operating-system component. Equally, assuming every Linux host is exposed would disregard version-specific fixes and configurations.

Temporary configuration changes have been discussed in the researcher’s advisory, but they can interfere with normal network performance and other protection mechanisms. One proposed trade-off weakens a defence against certain floods of connection requests. These are not neutral settings to apply indiscriminately to internet-facing systems. Prioritising vendor-supported corrections and testing their effect on affected services gives organisations a clearer route than treating an emergency configuration change as a substitute for a fix.

Kernel remediation can require operational coordination because installing an updated package may leave the previous kernel active until the machine restarts. A rolling restart plan is particularly relevant in clustered services where several hosts share the affected build. The timing of maintenance must account for continuity and the possibility that a newly updated node has not yet booted the corrected kernel. Monitoring can help identify disruptive symptoms, but successful detection after a crash does not eliminate availability risk.

NoisyTCP (CVE-2026-99288) has a separate identifier and fix history from CVE-2026-98432, another Linux TCP vulnerability disclosed earlier in November. The available advisory provides evidence of a pre-authentication crash condition and a reported upstream correction. It does not establish widespread attacks against production systems or a confirmed method of gaining privileged access. Further distribution-specific advisories will determine how the disclosed flaw translates into patch requirements for maintained enterprise installations.

×