high

CVE

CVE-2026-99288

CWE

CWE-362, CWE-416, CWE-476

Affected Surface

  • Linux kernel 4.4 and later (since commit `079096f103fa`, 3 October 2015) before 6.12.117, 6.18.63 and 7.3.4; no upstream fix yet in the 6.6, 6.1, 5.15 and 5.10 longterm lines
  • Any host with a reachable TCP listener and two or more RX queues, or RPS spreading packets across CPUs, including containers that share the host kernel
  • Debian bookworm before `6.1.192-1` and trixie before `6.12.117-1`; Ubuntu (all supported releases, needs triage); RHEL 8, 9 and 10 and OpenShift 4 and 5 (affected, no fix)

TL;DR

NoisyTCP (CVE-2026-99288) is a race in the Linux kernel’s lockless lookup of established TCP sockets. An attacker who can send packets to any TCP listener can make the receive path dereference a request socket whose listener pointer is NULL, which panics the kernel. Nothing has to be configured specially. The reporter, Hyunwoo Kim, triggered it against Nginx, Apache and OpenSSH running with default settings. The bug came in with Linux 4.4 in 2015.

The race needs two receive contexts processing packets for the same connection tuple at once. That means two or more NIC RX queues, or RPS spreading packets across CPUs. Most cloud VMs and servers meet that condition. On QEMU/KVM guests Kim needed 5 seconds to 2 minutes. On a quiet bare-metal Nginx box with 8 RX queues it took about 14 hours. A second path leads to a use-after-free, which Kim describes as hard to hit without special techniques. Red Hat rates the CVE Important (CVSS 7.5) and says it does not allow code execution.

Kim’s repository has the write-up but no exploit code. EPSS on 15 November was 0.0016. There are no reports of exploitation in the wild.

Containers don’t isolate you from this. A crash in the host kernel takes down every workload on the node.

What to do now: Update to 6.12.117, 6.18.63, 7.3.4 or a distribution kernel that includes commit 4e2b3ef19e0e. If you can’t, the advisory gives two stopgaps: set net.ipv4.tcp_syncookies=0 in every network namespace, or reduce the internet-facing interface to one RX queue with RPS and RFS off. Both have costs, covered below.

Affected packages and projects

EcosystemPackageAffected versionsFixed or safe version
Linux (upstream)kernel4.4 (commit 079096f103fa) and later6.12.117 (d69100340b4c), 6.18.63 (8e5172cbd227), 7.3.4 (eb8a2d156f2c), 7.4-rc2 (4e2b3ef19e0e)
Linux (upstream LTS)kernel 6.6, 6.1, 5.15, 5.10All releases up to 6.6.162, 6.1.192, 5.15.225, 5.10.274No upstream backport listed by the kernel CNA as of 16 November 2026
Debianlinuxbookworm before 6.1.192-1, trixie before 6.12.117-1, forky 7.2.15-1bookworm 6.1.192-1 (DLA-4906-1), trixie 6.12.117-1 (DSA-6628-1), sid 7.3.4-1
Ubuntulinux and variantsAll supported releasesNeeds triage, no fixed kernel yet
Red HatkernelRHEL 8, 9, 10 and OpenShift Container Platform 4 and 5Affected, no fixed kernel yet. RHEL 6 and 7 are not affected

Debian’s bookworm kernel is fixed even though the upstream 6.1 line has no backport, because Debian carries the patch itself.

Technical analysis

The bug

TCP looks up established sockets in the ehash table without taking a lock. The socket caches are created with SLAB_TYPESAFE_BY_RCU, so a freed socket’s memory can be handed out again while an RCU reader still holds the old address. Readers have to defend against that, and __inet_lookup_established() tries to: after a tuple match it takes a reference with refcount_inc_not_zero() and checks the tuple again.

That second check isn’t enough. If the slab slot was reused for a new socket with the same 4-tuple, the tuple still matches and the refcount is non-zero, so the lookup returns the new object. Kim’s PoC gets that new object to be a SYN cookie request socket built by tcp_get_cookie_sock(). That socket was never inserted into ehash and its rsk_listener is NULL. tcp_v4_rcv() then reads sk_state through the null listener:

BUG: kernel NULL pointer dereference, address: 0000000000000012
RIP: 0010:tcp_v4_rcv+0x44e/0x1280
Call Trace:
 <IRQ>
 ip_protocol_deliver_rcu+0x2d/0x180
 ip_local_deliver_finish+0x84/0x100
 ...
 virtnet_poll+0x445/0x600
 net_rx_action+0x2fe/0x3b0
Kernel panic - not syncing: Fatal exception in interrupt

The crash happens in softirq context, so the whole machine goes down, not just the listening process.

To get two RX contexts racing on one tuple, the advisory relies on IPv4 fragments and regular TCP segments getting different receive hashes, so they land on different queues or CPUs. Because the attacker only needs to reach a listening port, the authentication step of the service (SSH, HTTP) never comes into play.

The fix

Commit 4e2b3ef19e0e (“inet: reject unhashed sockets during established lookups”, reviewed by Eric Dumazet and Kuniyuki Iwashima) adds an “is it still hashed?” test after the reference is taken, for both IPv4 and IPv6:

--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -552,8 +552,9 @@ struct sock *__inet_lookup_established(const struct net *net,
 		if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
 			if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 				goto out;
-			if (unlikely(!inet_match(net, sk, acookie,
-						 ports, dif, sdif))) {
+			if (unlikely(hlist_nulls_unhashed_lockless(&sk->sk_nulls_node) ||
+				     !inet_match(net, sk, acookie, ports, dif,
+						 sdif))) {
 				sock_gen_put(sk);
 				goto begin;
 			}
  • refcount_inc_not_zero() still pins whatever object now lives at that address.
  • hlist_nulls_unhashed_lockless() is the new check. A SYN cookie request that was never put in ehash, or a socket that has left it, fails here.
  • On failure the lookup drops the reference and restarts from begin, so it finds the current socket, such as the child created by the TIME_WAIT hashdance or by request-to-child promotion.

__inet6_lookup_established() in net/ipv6/inet6_hashtables.c gets the same change.

Once a stale lookup can hold a short-lived reference to a request socket that was never hashed, freeing those requests directly is no longer safe. The rest of the patch swaps __reqsk_free() for reqsk_put() on the failure paths that had already set a non-zero refcount:

--- a/net/ipv4/syncookies.c
+++ b/net/ipv4/syncookies.c
@@ -218,6 +218,8 @@ struct sock *tcp_get_cookie_sock(struct sock *sk, struct sk_buff *skb,
 		bh_unlock_sock(child);
 		sock_put(child);
+		reqsk_put(req);
+		return NULL;
 	}
 	__reqsk_free(req);

The same swap covers the Fast Open path in tcp_conn_request() (net/ipv4/tcp_input.c) and three request-migration paths in net/ipv4/inet_connection_sock.c (reqsk_timer_handler(), inet_csk_complete_hashdance() and inet_csk_listen_stop()).

IOCs

This is a kernel vulnerability, not a malware campaign, so there are no domains, IPs or hashes. What an attempt looks like on the host:

  • A kernel panic or oops with RIP: tcp_v4_rcv (or the IPv6 equivalent) and BUG: kernel NULL pointer dereference inside <IRQ> / net_rx_action frames.
  • Before the crash, per the trigger the advisory describes, a burst of incoming IPv4 fragments mixed with TCP segments for one tuple aimed at a listening port, with SYN cookies being issued (TcpExtSyncookiesSent going up in nstat).

Detection and remediation

# 1. Is the running kernel in a fixed upstream release?
uname -r    # fixed upstream: >= 6.12.117 (6.12.y), >= 6.18.63 (6.18.y), >= 7.3.4 (7.3.y), 7.4-rc2+

# 2. Distribution kernels: check the changelog for the CVE or the fix title
apt changelog "linux-image-$(uname -r)" 2>/dev/null | grep -m1 -E 'CVE-2026-99288|reject unhashed sockets'
rpm -q --changelog kernel 2>/dev/null | grep -m1 -E 'CVE-2026-99288|reject unhashed sockets'

# 3. Do you meet the race preconditions? (more than one RX queue, or RPS enabled)
ethtool -l eth0 | sed -n '/Current hardware settings/,$p'
cat /sys/class/net/eth0/queues/rx-*/rps_cpus

# 4. Look for past crashes that match the signature
journalctl -k -b -1 2>/dev/null | grep -E 'tcp_v4_rcv|tcp_v6_rcv' | head
nstat -az TcpExtSyncookiesSent TcpExtSyncookiesRecv

If you can’t patch yet, the advisory gives two temporary mitigations. Each one costs you something:

# Option A: disable SYN cookies. Per network namespace, so repeat inside every container netns.
sysctl -w net.ipv4.tcp_syncookies=0
# Turning SYN cookies off weakens protection against SYN floods.

# Option B: serialize receive processing (one RX queue, RPS and RFS off). This costs throughput.
ethtool -L eth0 combined 1
for q in /sys/class/net/eth0/queues/rx-*; do echo 0 > "$q/rps_cpus"; echo 0 > "$q/rps_flow_cnt"; done
sysctl -w net.core.rps_sock_flow_entries=0

Patch order: start with internet-facing hosts that run many tenants or containers, such as Kubernetes nodes, shared CI runners and ingress or load-balancer hosts, because one crash takes down everything on them. No credentials are exposed, so nothing needs rotating.

Timeline

  • 2015-10-03: Commit 079096f103fa (“tcp/dccp: install syn_recv requests into ehash table”) introduces the bug. It ships in Linux 4.4.
  • 2026-10-29: Hyunwoo Kim posts the fix to netdev.
  • 2026-11-03: Paolo Abeni merges 4e2b3ef19e0e into the networking tree. It is released in 7.4-rc2.
  • 2026-11-10: Stable releases 6.12.117, 6.18.63 and 7.3.4 ship with the backport, and the kernel CNA publishes CVE-2026-99288 on linux-cve-announce.
  • 2026-11-11: Debian accepts linux 7.3.4-1 into unstable. Ubuntu lists the CVE as needs triage.
  • 2026-11-15: The linux-distros embargo ends. Kim posts the NoisyTCP advisory to oss-security and publishes the V4bel/NoisyTCP repository (16:33 UTC). By this date Debian has fixed kernels for bookworm (DLA-4906-1) and trixie (DSA-6628-1).
  • 2026-11-16: Red Hat lists RHEL 8, 9 and 10 as affected with no fixed kernel, and Ubuntu still has no fixed kernel.

From research to remediation

Check whether this pattern exists in your codebase

Turn this research into a remediation workflow. Scan dependencies and package manifests for similar supply-chain risk, then prioritize fixes with reachability context.

References