
A newly disclosed Linux kernel flaw could let an unauthenticated attacker crash a vulnerable server over the network. Security researcher Hyunwoo Kim calls the bug NoisyTCP and tracks it as CVE-2026-99288. The issue sits in the kernel’s TCP handling, so an ordinary exposed service can provide a route to the fault before a user logs in.
Kim said default configurations of Nginx, Apache and SSH were affected in his tests. The headline risk demonstrated publicly is a kernel panic and denial of service, not confirmed remote code execution. His advisory also describes a harder-to-trigger use-after-free path, but it does not establish a general-purpose exploit for taking control of servers.
How NoisyTCP works
The vulnerability is a race in the Linux TCP stack. Processing of crafted incoming traffic can be split between different receive contexts, leaving a stale socket reference at precisely the wrong moment. Kim said the condition requires at least two receive queues or software that distributes packet processing across multiple CPUs. That makes the risk dependent on a server’s kernel and networking setup, even though the underlying code has been widespread.
In his virtual-machine tests, triggering the crash against common services took about five seconds to two minutes. A quiet physical server took far longer in a tuned test. Those timings show both why the bug matters for internet-facing infrastructure and why it would be misleading to suggest every Linux machine will fail immediately on contact.
Kim traces the vulnerable code to an October 2015 change and identifies the upstream fix as commit 4e2b3ef19e0e, merged on 3 November 2026. Linux distributions package their own kernels and backport fixes, so administrators should consult their vendor’s advisory rather than rely on a broad version number alone. They also need to confirm the patched kernel is the one actually running after an update and restart.
The flaw is particularly relevant to hosts exposing TCP services, including web and remote-access servers. A container does not automatically contain a kernel crash, because containers on the same host share the kernel. Organisations should prioritise exposed systems and assess whether their networking configuration meets the conditions described in the research.
Patch first, then consider mitigations
Applying a vendor-provided kernel fix is the preferred response. Kim describes temporary measures for systems that cannot be patched immediately, including reducing an external interface to one receive queue or disabling SYN cookies. Those changes carry operational and security trade-offs. In particular, disabling SYN cookies weakens a defence against SYN-flood attacks, so it should not be treated as a casual one-line fix.
The disclosure follows a run of high-profile software security findings, but NoisyTCP is a distinct kernel issue. Its practical lesson is familiar: internet-facing infrastructure needs prompt patching, a clear inventory of running kernels and careful testing of any emergency mitigation.







