← Voltar ao feed
high

NoisyTCP (CVE-2026-99288): Linux kernel TCP race enables remote pre-auth kernel panic, fixed in 6.12.117, 6.18.63 and 7.3.4

2 fonte(s) · 2 atualização(ões) · publicado 15/11/2026 22:35 · coletado há 25 minutos

CVE-2026-99288, named NoisyTCP by researcher Hyunwoo Kim, is a race condition in the Linux kernel TCP established-connection lookup. It has existed since commit 079096f103fa of October 3, 2015 (gbhackers.com). A remote, unauthenticated attacker can send crafted TCP traffic to trigger a kernel panic. A second path can lead to a use-after-free, but Kim calls it hard to trigger. No exploitation in the wild has been reported, and the flaw is not in CISA KEV. No public exploit code exists, and Kim says it is not a TCP RCE so far. Many distributions had not shipped patched kernels as of November 15 (gbhackers.com). This summary is based on the articles alone. No vendor or primary-source research was done, so no CVSS score is available.

Entenda a falha

A vulnerabilidade

CVE-2026-99288, named NoisyTCP, is a race condition in the Linux kernel TCP established-connection lookup. A remote, unauthenticated attacker who can reach an ordinary TCP listener can send crafted traffic and cause a kernel panic. A second path could lead to a use-after-free, but it needs nested races and is hard to trigger. The race needs two or more RX queues, or RPS spreading packets across CPUs, so default Nginx, Apache and SSH servers are exposed. Affected: Linux kernel 4.4 up to before 6.12.117, and the 6.18 series before 6.18.63. Fixed in 6.12.117, 6.18.63, 7.3.4 and 7.4-rc2. No CVSS score is available.

Como funciona

TCP looks up established sockets without locks, using RCU. One CPU can hold a stale pointer to a request socket while another RX context frees it and reallocates a SYN cookie request with the same tuple at the same slab address. The stale lookup then points at the new object and passes the tuple recheck, but nothing confirms the socket is hashed. The new request has no attached listener, so the TCP receive path dereferences a NULL listener and the kernel panics. On a container host the crash takes down the entire host, not only the web server.

Por que é vulnerável

Socket memory can be freed and reused while an RCU reader still holds its old address. The lookup rechecks the connection tuple but does not verify that the socket is hashed, so it trusts memory that now holds a different object. The fix, commit 4e2b3ef19e0e (inet: reject unhashed sockets during established lookups), makes established lookups reject unhashed sockets. The bug was introduced by commit 079096f103fa on October 3, 2015 and went unnoticed for about 11 years, like other long-lived object lifetime errors.

Onde está

The fix is commit 4e2b3ef19e0e (inet: reject unhashed sockets during established lookups), which landed on November 3, 2026, in the kernel TCP established lookup; the flaw was introduced by commit 079096f103fa.

Como foi descoberta

Researcher Hyunwoo Kim (@v4bel) found the bug, named it NoisyTCP and, according to gbhackers, also patched it. After the agreed embargo with Linux distributions ended, he published the full advisory on the oss-security mailing list. He triggered it in about 5 seconds to 2 minutes on QEMU/KVM VMs, where vCPU preemption widens the race window, and in about 14 hours on a quiet bare-metal Nginx server with 8 RX queues. The articles do not say how he first found it. No exploitation in the wild has been reported and the flaw is not in CISA KEV. As of November 15, many distributions had not shipped patched kernels.

PoC existente, privada

O exploit existe, mas quem o tem não o publicou.

A PoC exists but is not public. Kim used one with carefully tuned timing to crash a quiet bare-metal Nginx server in about 14 hours, and a lab-specific exploit against VM servers, but no exploit code has been released. Public reports document only a denial-of-service crash; neither Kim nor vendors have indicated remote code execution is possible, and he says it is not a TCP RCE so far. Since the crash is reproducible in seconds on VMs, patching remains urgent. If you cannot patch, set net.ipv4.tcp_syncookies to 0 in every network namespace (this weakens SYN flood protection) or limit external interfaces to one RX queue and disable RPS and RFS.

Explicação gerada por IA a partir da matéria e dos artigos, em 16/11/2026 05:16.

Detalhes técnicos

The lockless RCU lookup of established sockets can see socket memory that has been freed and reused. The kernel rechecks the connection tuple but does not check that the new socket is hashed. A request socket is freed and its memory is reused for a SYN cookie request at the same slab address. That request has no listener, so the kernel dereferences a NULL pointer and panics. The race needs two or more RX queues, or RPS spreading packets across CPUs. It triggers even when a server only calls listen(), so default Nginx, Apache and SSH are exposed. Kim triggered it in about 5 seconds to 2 minutes on QEMU/KVM VMs, and in about 14 hours on a quiet bare-metal Nginx server with 8 RX queues (gbhackers.com). A crash on a container host takes down the entire host. The second scenario recycles a full child socket and needs nested races (gbhackers.com). The affected range is Linux kernel 4.4 and later, before 6.12.117, and the 6.18 series before 6.18.63.

Remediação

Install a kernel carrying commit 4e2b3ef19e0e ("inet: reject unhashed sockets during established lookups"), which landed on November 3, 2026 (gbhackers.com), or your distribution's backport. Fixed releases are 6.12.117, 6.18.63, 7.3.4 and 7.4-rc2. Reboot into the patched kernel and confirm it with uname -r (gbhackers.com). Two stopgaps exist. The first is to set net.ipv4.tcp_syncookies to 0 in every network namespace, including containers, which weakens SYN flood protection. The second is to limit external interfaces to a single RX queue and disable RPS and RFS.

Linha do tempo

  1. gbhackers.com ↗ · pub. 16/11/2026 04:51 · coletado 16/11 05:00

    NoisyTCP (CVE-2026-99288): 11-year-old Linux kernel TCP race lets remote attackers panic servers without authentication

    The article adds the introducing commit 079096f103fa (October 3, 2015), the fix date of November 3, 2026, the 8-RX-queue bare-metal test, and the note that many distributions had not shipped patched kernels as of November 15. It also adds the reboot-and-verify advice.

    ler a análise desta fonte

    The article adds the introducing commit 079096f103fa (October 3, 2015), the fix date of November 3, 2026, the 8-RX-queue bare-metal test, and the note that many distributions had not shipped patched kernels as of November 15. It also adds the reboot-and-verify advice.

    A Linux kernel TCP race condition, NoisyTCP (CVE-2026-99288), was found and patched by researcher Hyunwoo Kim (@v4bel). It lets a remote, unauthenticated attacker crash a server. The flaw dates to commit 079096f103fa of October 3, 2015. Kim published the full advisory on oss-security after the distribution embargo ended. The article says only a denial-of-service crash is publicly documented, and neither Kim nor any vendor has indicated that RCE is possible. Many distributions had not shipped patched kernels as of November 15. This summary rests on the article alone. No primary-source research was done.

    Technical Details

    A stale established-hash (ehash) lookup races with the freeing and reallocation of a request socket in a different RX context. In the first scenario the request is reallocated as a SYN cookie request at the same slab address with the same tuple. It passes the tuple recheck but has no attached listener, so the kernel dereferences a NULL listener and panics. This crashes the whole host, including container hosts. The second scenario recycles a full child socket and could lead to a remote use-after-free. It needs nested races and is hard to trigger. The race needs two or more RX queues, or RPS. Default Nginx, Apache and SSH listeners are exposed. In QEMU/KVM VMs, crashes took about 5 seconds to 2 minutes. A bare-metal Nginx server with 8 RX queues took about 14 hours.

    Fix Details

    The fix is commit 4e2b3ef19e0e, which landed on November 3, 2026. Upstream fixes are in Linux 6.12.117, 6.18.63 and 7.3.4.

    Remediation

    Install the patched kernel, reboot, and confirm it with uname -r. If patching is not possible, there are two stopgaps. The first is to set net.ipv4.tcp_syncookies=0 in each network namespace, which weakens SYN flood protection. The second is to limit the inbound interface to a single RX queue and disable RPS and RFS.

  2. securityonline.info ↗ · pub. 15/11/2026 22:35 · coletado 15/11 23:00 primeiro registro

    NoisyTCP (CVE-2026-99288): Linux kernel TCP race enables remote pre-auth kernel panic, fixed in 6.12.117, 6.18.63 and 7.3.4

    ler a análise desta fonte

    CVE-2026-99288, named NoisyTCP by researcher Hyunwoo Kim, is a race condition in the Linux kernel TCP established-connection lookup. It has existed since a 2015 commit. A remote, unauthenticated attacker can send crafted TCP traffic to trigger a kernel panic. A second path can lead to a use-after-free, but Kim calls it hard to trigger. No exploitation in the wild has been reported, and the flaw is not in CISA KEV. No public exploit code exists, and Kim says it is not a TCP RCE so far. This summary is based on the article alone. No vendor or primary-source research was done, so no CVSS score is available.

    Technical Details

    The lockless RCU lookup of established sockets can see socket memory that has been freed and reused. The kernel rechecks the connection tuple but does not check that the new socket is hashed. A request socket is freed and its memory is reused for a SYN cookie request. That request has no listener, so the kernel dereferences a NULL pointer and panics. The race needs two or more RX queues, or RPS spreading packets across CPUs. It triggers even when a server only calls listen(), so default Nginx, Apache and SSH are exposed. Kim triggered it in about 5 seconds to 2 minutes on QEMU/KVM VMs, and in about 14 hours on a quiet bare-metal Nginx server. A crash on a container host takes down the entire host. The affected range is Linux kernel 4.4 and later, before 6.12.117, and the 6.18 series before 6.18.63.

    Remediation

    Install a kernel carrying commit 4e2b3ef19e0e ("inet: reject unhashed sockets during established lookups"), or your distribution's backport. Fixed releases are 6.12.117, 6.18.63, 7.3.4 and 7.4-rc2. Two stopgaps exist. The first is to set net.ipv4.tcp_syncookies to 0 in every network namespace, including containers, which weakens SYN flood protection. The second is to limit external interfaces to a single RX queue and disable RPS and RFS.