News Security news
NoisyTCP (CVE-2026-99288): a Linux TCP race that crashes servers remotely
On 15 November 2026 researcher Hyunwoo Kim published details of NoisyTCP, a Linux kernel flaw. A race in the TCP connection lookup lets an unauthenticated attacker trigger a kernel panic on any server that simply listens on a TCP port. The bug dates back to 2015, and fixed kernels are already out.
Key points
- CVE-2026-99288 (NoisyTCP) is a race condition in the Linux kernel's TCP socket lookup. The lockless lookup can return a socket that has been freed and reused, and the kernel then dereferences a NULL pointer in tcp_v4_rcv(), crashing the machine.
- The researcher says it can be triggered remotely and without authentication against default setups of nginx, Apache and SSH. It requires SYN cookies (on by default) and more than one receive context – two or more RX queues, or RPS. A panic inside a container takes down the whole host.
- The bug was introduced by a commit dated 3 October 2015 (kernel 4.4). It is fixed in 7.4-rc2 and in stable kernels 6.12.117, 6.18.63 and 7.3.4. Debian shipped fixes in DSA-6628-1 (trixie) and DLA-4906-1 (bookworm); Red Hat lists RHEL 8, 9 and 10 as affected.
- As a stopgap, disable SYN cookies (net.ipv4.tcp_syncookies=0, which weakens SYN flood protection) or restrict the internet-facing interface to one receive queue with RPS and RFS off.
Vulnerabilities: CVE-2026-99288
What happened
On 15 November 2026 researcher Hyunwoo Kim (@v4bel) published details on oss-security and GitHub of a bug named NoisyTCP (CVE-2026-99288). The bug had been reported privately to distribution maintainers through the linux-distros list, and the details went public once the agreed embargo ended.
The flaw sits in the Linux kernel's TCP stack. According to the researcher, it lets an attacker crash the kernel remotely and without authentication (a kernel panic) on any server that simply listens on a TCP port. The researcher reproduced it against nginx, Apache and SSH in their default configurations. The NVD record went live on 10 November 2026, and the fix is in mainline (commit 4e2b3ef19e0e) and the stable branches.
The bug has been in the kernel for roughly 11 years: it was introduced by commit 079096f103fa on 3 October 2015, which shipped in kernel 4.4.
Technical details
| Item | Value |
|---|---|
| Component | established TCP connection hash table lookup (ehash), inet |
| Bug class | race condition (ABA) → NULL pointer dereference in tcp_v4_rcv() |
| Authentication | none required |
| Preconditions | SYN cookies enabled (default), two or more RX queues or RPS |
| Impact | kernel panic (DoS); a second, harder path leads to use-after-free |
| CVSS | not yet scored by NVD; Red Hat: Important, draft 7.5 |
The kernel looks up which socket an incoming TCP packet belongs to without taking locks. According to the fix description in NVD, the lookup takes a reference on the object it finds but does not check whether that socket is still hashed. In the meantime the original connection request (request socket) can be freed and its memory reused for a new object created via SYN cookies. That object has no listener attached, and when the kernel carries on processing the packet it dereferences a NULL pointer.
For the race to happen, packets for the same connection have to be handled by different receive contexts at the same time. The researcher achieves this, according to the write-up, by exploiting the difference in how the receive hash is calculated for IPv4 fragments versus ordinary TCP packets. On QEMU/KVM virtual machines the crash took between 5 seconds and 2 minutes; on a quiet bare-metal server with 8 queues it took about 14 hours. We don't go further into the attack itself.
The fix ("inet: reject unhashed sockets during established lookups") checks, after taking the reference, that the socket is hashed and still matches the connection, and restarts the lookup if it doesn't.
Who is affected
Unpatched Linux kernels from 4.4 onwards on systems that listen on a TCP port and accept packets from the internet or another untrusted network. The researcher says no special configuration is needed beyond calling listen(). A panic triggered through a container brings down the entire host and every container on it.
As of 16 November NVD lists no fixed version for other stable branches (such as 6.6, 6.1, 5.15 and 5.10), so watch your distribution's security advisories. Red Hat says RHEL 6 and 7 are not affected, and also lists RHEL CoreOS in OpenShift 4 and 5 as affected.
How to tell if you were compromised
Neither the researcher nor the distributions have published indicators of compromise, and nobody has reported exploitation in the wild. A successful attack shows up as a server crash and reboot. Based on the bug description, this is what we would look for:
- Unexpected reboots of servers running internet-facing TCP services.
journalctl -k -b -1shows the kernel log from the previous boot (if the journal is persistent); kdump or netconsole can also capture it. - In the crash output, look for "BUG: kernel NULL pointer dereference" with
tcp_v4_rcvin the call trace. - In flow or firewall logs, an unusual volume of fragmented IPv4 packets alongside SYNs to listening ports from the same source.
You can pull source addresses out of a log export with our IOC extractor, which runs entirely in your browser.
What to do
- Install the fixed kernel from your distribution's repositories and reboot. Check the running version with
uname -r. - Prioritise servers directly reachable from the internet and container hosts.
- As a stopgap, if no fix is available yet, apply one of the measures recommended by the researcher and Red Hat:
- disable SYN cookies:
sysctl -w net.ipv4.tcp_syncookies=0– this is per network namespace and weakens SYN flood protection, - or limit the internet-facing interface to a single receive queue (check with
ethtool -l <interface>) and disable RPS and RFS – this may cost performance.
- disable SYN cookies:
- Once patched, revert the stopgap (
net.ipv4.tcp_syncookies=1).
Timeline
Updates
- 16 November 2026 – first version of this article.
FAQ
Can NoisyTCP be used to take over a server?
Not on current information. Red Hat and the researcher describe the practical impact as a kernel crash (denial of service). The researcher mentions a second, harder-to-trigger path that leads to a use-after-free, but says no general-purpose code execution exploit exists yet. That is still no reason to delay patching.
Is restarting nginx or sshd enough?
No. The bug is in the kernel, not the application. Only a fixed kernel and a reboot help, or a live patch if your distribution ships one for this CVE.
We run containers. Do we just rebuild our images?
No. Containers share the host kernel, so the host needs the fix. If you disable SYN cookies as a stopgap, remember the setting is per network namespace, so it also has to be applied inside every container with its own network namespace.
Is it safe to turn off SYN cookies?
It is a trade-off. SYN cookies protect a server during a SYN flood, and without them it is more exposed to that kind of attack. Treat it as a temporary measure until the fixed kernel is installed.
Sources
- Hyunwoo Kim – NoisyTCP: Linux Kernel TCP Pre-auth Remote Corruption (CVE-2026-99288)
- oss-security – NoisyTCP: Linux Kernel TCP Pre-auth Remote Corruption (15 November 2026)
- NVD – CVE-2026-99288
- Linux kernel – commit 4e2b3ef19e0e: inet: reject unhashed sockets during established lookups
- Red Hat – CVE-2026-99288
- Debian Security Tracker – CVE-2026-99288
Spotted a mistake? Let us know and we will correct it openly.