TL;DR
A race condition in Linux kernel TCP code, tracked as CVE-2026-99288, can let a remote attacker crash a server before any login. Researcher Hyunwoo Kim named it the NoisyTCP vulnerability. It has sat in the kernel since a 2015 commit, roughly 11 years.
The NoisyTCP vulnerability lets crafted TCP traffic trigger a kernel panic on servers with multiple receive queues. A second path can lead to a use-after-free, though Kim calls it hard to trigger. Fixes landed in kernels 6.12.117, 6.18.63, 7.3.4 and 7.4-rc2.
- CVE: CVE-2026-99288
- CVSS: Awaiting analysis
- Product: Linux
- Affected: 079096f103faca2dd87342cca6f23d4b34da8871, 4.4
- Impact: inet: reject unhashed sockets during established lookups
- Status: No confirmed exploitation yet
- Patched in: d69100340b4cae08d40f898f1b601032403f6be6, 8e5172cbd227fae32f3313cba1f3e1b9edb5c4fd, eb8a2d156f2c1d862fffdfd1b3432bfbcd939879, 4e2b3ef19e0e04adca38442f07dd6bc93c712bf1 (+4 more)
- EPSS: 0.2% (30-day)
- Action: Update to d69100340b4cae08d40f898f1b601032403f6be6, 8e5172cbd227fae32f3313cba1f3e1b9edb5c4fd, eb8a2d156f2c1d862fffdfd1b3432bfbcd939879, 4e2b3ef19e0e04adca38442f07dd6bc93c712bf1 (+4 more) now
Why It Matters
The bug needs no special setup. According to Kim’s NoisyTCP advisory, it triggers “even when the server does nothing more than call listen().” That covers Nginx, Apache and SSH servers in their default configuration.
The impact also reaches past one app. Kim warns that a crash in a containerized web server “takes down not only the web server application but the entire host.”
How the Attack Works
A Stale Socket Lookup
TCP looks up established connections without locks. The kernel CVE record explains that socket memory “can be reused before an RCU reader that retained its old address has finished.” The lookup then checks the connection details again, but it never checks whether the new socket is actually hashed.
An attacker can race this window. One CPU holds a stale lookup. Meanwhile, another frees the request socket and reuses its memory for a SYN cookie request. That new request has no listener, so the kernel hits a NULL pointer and panics.
Timing Limits
The race needs two or more RX queues, or RPS spreading packets across CPUs. On QEMU/KVM virtual machines, Kim triggered it in about 5 seconds to 2 minutes. On a quiet bare-metal Nginx server, however, it took about 14 hours.
Exploitation Status
Kim has not released exploit code. The repository holds only the write-up. He says he wrote only “a very limited, lab-specific exploit that targets VM servers.” He also says the bug is not a TCP RCE so far. No attacks in the wild have been reported, and the flaw is not in CISA’s KEV catalog.
Affected Versions
- Linux kernel 4.4 and later, before 6.12.117
- 6.18 series before 6.18.63
- 7.3 series before 7.3.4
Patch and Mitigation
Install a kernel with commit 4e2b3ef19e0e, or your distribution’s backport. Check the Red Hat CVE page and other vendor trackers for package status.
If you cannot patch yet, Kim offers two stopgaps. First, set net.ipv4.tcp_syncookies to 0 in every network namespace, including containers. This weakens SYN flood protection, though. Alternatively, cut external interfaces to a single RX queue and disable RPS and RFS. Either step closes the NoisyTCP vulnerability race until the fix is in place.
Support Our Threat Intelligence
Find our vulnerability reports and weekly recaps helpful? Support our work today and unlock a 100% ad-free reading experience!