NoisyTCP: an 11-year-old Linux kernel bug can crash servers remotely
CVE-2026-99288 is a race condition in the kernel TCP stack dating back to a 2015 commit. No login is needed, and it hits any server that just calls listen().
Researcher Hyunwoo Kim disclosed NoisyTCP, tracked as CVE-2026-99288, on the oss-security mailing list on 15 November. It’s a race condition in the Linux kernel’s TCP code: a stale lookup in the connection hash table can collide with a request socket being freed and reallocated in a different RX context, leading to a null pointer dereference or, in the full child-socket path, a use-after-free. The bug traces back to a commit from October 2015, roughly 11 years ago.
It triggers on any server with more than one RX queue, or with Receive Packet Steering spreading packets across CPUs, which covers most multi-core machines by default. No authentication is needed, just correctly timed TCP packets, and it works against anything that calls listen(), including Nginx, Apache and SSH in their stock configuration. Kim’s write-up notes that on a container host, the crash takes down the whole machine, not just the one service. The fix merged upstream on 3 November, landing in 6.12.117, 6.18.63 and 7.3.4. Red Hat rates it 7.5, Important.
Why it matters: this is a pre-authentication denial-of-service sitting in kernel code that almost every Linux box already runs. Check your distribution has shipped a kernel built past the fix commit before assuming you’re covered; sysctl net.ipv4.tcp_syncookies=0 or pinning traffic to a single RX queue are stopgaps Red Hat lists, not fixes.
The caveat: no exploit code is public, it isn’t in CISA’s KEV catalogue, and nobody has reported it being used in the wild. This is a serious, disclosed bug in widely deployed infrastructure, not yet a confirmed attack.