The distro embargo for NoisyTCP ended on November 15, 2026. Researcher Hyunwoo Kim (@v4bel), who reported the bug and wrote the fix, published the details on oss-security. The bug is a remote, unauthenticated kernel panic in the Linux TCP receive path. It was introduced in October 2015 (Linux 4.4), so almost every distribution is affected. The fix, 4e2b3ef19e0e, was merged upstream on November 3 and is now in the stable trees and the first distro kernels.
What goes wrong
When a TCP segment arrives, the kernel looks up its socket in the established hash table (ehash) without taking a lock. The socket caches use SLAB_TYPESAFE_BY_RCU, which means a socket’s memory can be freed and reused while a reader on another CPU still holds the old address. The lookup guards against this by taking a reference and checking the address/port tuple again. According to the kernel’s CVE text, it never checked whether the object it now holds is actually in the hash table.
Kim’s proof of concept takes advantage of that:
- One RX context performs a lookup and finds an existing request socket.
- Another RX context frees that request socket and allocates a SYN-cookie request socket with the same tuple at the same slab address.
- The stale lookup takes a reference on the new object, and the tuple check passes.
- A SYN-cookie request socket isn’t hashed and has a NULL
rsk_listener.tcp_v4_rcv()readssk_statethrough it, and the result is “Kernel panic - not syncing: Fatal exception in interrupt.”
Kim describes a second trigger, a “full child socket” ABA race, that leads to a use-after-free. He calls it “generally hard to trigger without using special techniques”. Red Hat’s assessment is that the flaw affects availability and does not allow data changes or code execution.
Who is exposed
Any Linux host that receives TCP from the attacker directly and runs a listener: Nginx, Apache or an SSH server with default settings is enough. The race needs packets processed in different RX contexts at the same time. The attacker arranges that by exploiting the different receive hashes for IPv4 fragments and normal TCP segments, so the host needs two or more RX queues or RPS enabled. AlmaLinux notes that this “describes most servers and cloud instances”.
Kim’s timings:
| Target | Time to crash |
|---|---|
| Default Nginx / Apache / SSH on QEMU/KVM VMs | ~5 seconds to 2 minutes |
| Default Nginx on a quiet bare-metal server (8 RX queues), tuned PoC | ~14 hours |
The crash is kernel-wide. A web server in a container doesn’t contain it: the host and every container on it go down. VMs are easier targets than bare metal because vCPU preemption widens the race window.
Fixed versions
| Source | Fixed in |
|---|---|
| Upstream | 4e2b3ef19e0e “inet: reject unhashed sockets during established lookups” (7.4-rc2) |
| Stable | 6.12.117, 6.18.63, 7.3.4 |
| Debian | bookworm 6.1.192-1 (DLA-4906-1), trixie 6.12.117-1 (DSA-6628-1), sid 7.3.4-1 |
| AlmaLinux (testing) | 8: kernel-4.18.0-553.186.2.el8_10 · 9: kernel-5.14.0-741.22.2.el9_9 · 10: kernel-6.12.0-264.16.2.el10_3 |
| RHEL | No fixed kernel yet as of Nov 16. Red Hat rates it Important (draft CVSS 7.5) |
The patch rejects lookups that land on an unhashed object and restarts them. It also makes the SYN-cookie, Fast Open and request-migration error paths release request sockets with reqsk_put() instead of freeing them directly.
What to do
- Patch and reboot first, starting with internet-facing hosts and container nodes. On AlmaLinux, the fixed kernels are in
almalinux-testing(dnf update 'kernel*' --enablerepo=almalinux-testing). Turn that repo off again after the reboot. - If you can’t reboot yet, pick one stopgap and accept its cost:
- Disable SYN cookies with
sysctl -w net.ipv4.tcp_syncookies=0. The setting is per network namespace, so it has to be applied on the host and in every container that has its own netns. It removes your SYN-flood protection. AlmaLinux also notes that it doesn’t cover Kim’s second, harder variant. - Serialize receive processing with
ethtool -L <iface> combined 1(orrx 1) and keep RPS/RFS off. This costs throughput, doesn’t persist across reboots, and some virtual NICs don’t support it.
- Disable SYN cookies with
- Watch for unexplained kernel panics on TCP-facing hosts, especially VMs. A panic in
tcp_v4_rcvis the PoC’s crash signature.
This is the second serious TCP-stack bug in a month. Our earlier coverage of CVE-2026-98432 is a different flaw with a different fix, so applying one patch doesn’t cover the other.