Skip to content
audiotale.net
Sign in
NewsNews

CVE-2026-99288 NoisyTCP: remote Linux kernel crash via TCP race, fixed in Debian kernels

An 11-year-old race in Linux TCP socket lookup lets an unauthenticated attacker panic a multi-queue server, public proof-of-concept code exists, and patches are rolling out.

audiotale.net team

· 6 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What to do
  3. What happened
  4. Background
  5. How it works
  6. What it means
  7. What we don't know yet
  8. What to watch next
  9. Sources

A researcher has published a working crash for a remote Linux kernel bug that sat in the TCP code for about 11 years. Tracked as CVE-2026-99288 and nicknamed NoisyTCP, it lets an unauthenticated attacker who can send TCP traffic to a server panic the kernel and take the machine down, and in a container host that means every container on it. The fix is in the mainline kernel and in Debian's kernels, and public proof-of-concept code exists, so patching should not wait. The attack is not easy to land, though: it needs a particular network setup and, on quiet bare metal, a very long time.

Key points

  • CVE-2026-99288 ("NoisyTCP") is a race condition in the Linux kernel's lockless TCP socket lookup that ends in a NULL pointer dereference and a kernel panic, found by Hyunwoo Kim.
  • The researcher's write-up says the flawed code was introduced in commit 079096f103fa on 3 October 2015 and fixed in commit 4e2b3ef19e0e, listed as 3 November 2026.
  • It is a remote, pre-authentication denial of service. Code execution has not been demonstrated; the author says only a very limited, lab-specific exploit exists for virtual machines.
  • The target needs two or more receive (RX) queues, or Receive Packet Steering (RPS) spreading packets over CPUs. Red Hat's preliminary score is CVSS 7.5 ("Important"); NVD had no score when we checked.
  • Debian lists fixed kernels: 6.1.192-2 (bookworm) and 6.12.117-1 (trixie). The workarounds are net.ipv4.tcp_syncookies=0 or a single RX queue with RPS and RFS off.

What to do

Who is affected. The write-up says the bug affects almost all Linux distributions, and that plain TCP listeners are enough: it names Nginx, Apache and SSH in default configurations as reachable. The researcher's page gives no list of affected kernel versions beyond the two commits (introduced 2015, fixed November 2026), so treat any kernel without the fix as vulnerable until your vendor says otherwise.

Patch. Install your distribution's kernel update and reboot. Debian's tracker shows these fixed versions:

  • Debian 12 (bookworm): fixed in 6.1.192-2, per DLA-4906-1 (vulnerable in 6.1.176-1).
  • Debian 13 (trixie): fixed in 6.12.117-1, per DSA-6628-1 (vulnerable in 6.12.107-1).
  • forky: fixed in sid's 7.3.4-1, not yet released to forky when we checked (vulnerable in 7.2.15-1).

Red Hat rates the flaw "Important" with a preliminary CVSS v3.1 score of 7.5 and links a bug entry and three upstream stable commits. Its page, as we read it, did not say whether an erratum was available, so check the Red Hat CVE page and your own vendor's tracker for Ubuntu, SUSE, Red Hat and others. We could not find an Ubuntu or SUSE advisory for this CVE in a search, which may only mean they are not published yet.

Workarounds if you cannot reboot yet. The researcher and Red Hat give the same two options:

  1. Disable SYN cookies with sysctl -w net.ipv4.tcp_syncookies=0. This blocks the path where a request socket is rebuilt from a SYN cookie. It is set per network namespace, so each container with its own namespace needs it, and it weakens protection against SYN floods.
  2. Reduce the internet-facing network interface to a single RX queue and turn off RPS and RFS on it. That removes the two concurrent receive contexts the race needs.

Check whether you are exposed. Run uname -r and compare it with your vendor's fixed version. Use ethtool -l <interface> to see how many queues the card uses, and look at /sys/class/net/<interface>/queues/rx-*/rps_cpus for RPS. These are standard Linux checks, not instructions from the advisories; a host with one queue and no RPS lacks the condition the write-up requires, but patching is still the fix.

What happened

BleepingComputer reported the flaw on 16 November 2026 and credited it to security researcher Hyunwoo Kim, who goes by @v4bel. The researcher published a write-up and proof-of-concept repository called NoisyTCP. According to the write-up, the embargo agreed after reporting to the linux-distros mailing list has ended. It does not give dates for the report, patch availability or disclosure, so we cannot say how long distributors had to prepare.

The mainline fix was merged about two weeks before the news coverage, according to BleepingComputer, which matches the commit date of 3 November in the write-up.

Background

Linux finds the socket a TCP packet belongs to by looking it up in a hash table of established connections. For speed, that lookup runs without taking a lock. Debian's description of the bug explains the consequence: after matching the connection's address and port tuple, the lookup took a reference to the socket without checking that it was still in the hash table. If the socket's memory had been freed and reused in the meantime, the lookup could return an object that was no longer hashed.

The researcher calls the result an "ABA race": the memory at one address is freed and a new request socket with the same tuple is allocated at the same address, so a stale lookup still matches. The new request socket has no listener attached. When the receive path then uses its listener pointer, which is NULL, the kernel panics.

The same researcher is behind other kernel findings this year, according to BleepingComputer: Dirty Frag in May 2026, a local privilege escalation chaining CVE-2026-43500 and CVE-2026-43284, and Januscape in July 2026, a 16-year-old flaw allowing a virtual machine to escape to its host.

How it works

The write-up describes two things that have to happen at the same time on different CPUs:

  • One receive context does a stale lookup in the established-connection table and finds the old request socket.
  • Another context frees that request socket and reallocates it at the same slab address with the same tuple.

The stale lookup takes a reference on the new object and passes the tuple recheck. Because the new object has no listener, the NULL dereference follows.

To make two CPUs process related packets at once, the attacker exploits how the receive hash differs between IPv4 fragments and regular TCP packets, which splits their processing across queues or CPUs. That is why the target needs two or more RX queues or RPS. The write-up also describes a second route, a race on a full child socket that could in principle lead to a remote use-after-free, but it says this is hard to trigger because nested races must succeed.

Timing. The race window is narrow. The write-up reports about 5 seconds to 2 minutes on QEMU/KVM virtual machines, and about 14 hours on a quiet bare-metal server with eight RX queues running Nginx. A busy production server generates its own traffic and may be easier to hit than a quiet test machine, but the write-up does not measure that, so we do not claim it.

What it means

For operators, the practical risk is a remote crash with no credentials. A server that takes TCP connections directly, on a multi-queue network card, is the profile in the write-up. Container hosts are the sharper case, because the panic is in the host kernel and stops every workload on the machine. Cloud load balancers and NAT in front of a server may hide the attacker's packets from it, but we have no source that tests this.

For defenders, the long wait for a reliable trigger matters in both directions: it limits mass exploitation, but public proof-of-concept code lowers the cost for anyone willing to run it for hours against a target they care about.

The score is also unsettled. Red Hat's vector rates confidentiality, integrity and availability impact as high, yet its own text says the flaw does not allow data alteration or code execution. NVD showed no score. Expect the numbers to change; plan by the practical impact, a remote crash.

What we don't know yet

  • Which stable kernel releases carry the fix for each long-term series, and when Ubuntu, SUSE, Red Hat and others ship it. We only confirmed Debian's versions.
  • Whether the "limited, lab-specific" virtual-machine exploit the author mentions can be generalised to code execution.
  • Whether cloud networking such as load balancers or virtual NICs prevents the packet pattern in real deployments.
  • Any evidence of attacks in the wild: none of the sources we read reported any.

What to watch next

Distribution advisories and kernel updates over the coming days, a final CVSS score from NVD, and any update to the NoisyTCP repository on exploitability. If you run containers or multi-queue network cards on internet-facing hosts, move this up your patch list.

Sources

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.