Incident sheet · SIG-6317 · /security
- Signal
- SIG-6317
- System
- SEC · Security & Risk
- Filed
- Read
- 3 min · 557 words
- Severity
- Routine
Eleven-year-old Linux TCP bug lets remote attackers crash default servers
NoisyTCP, a 7.5-rated kernel flaw present since October 2015, crashes default QEMU/KVM hosts in five seconds to two minutes. Fixed in kernels 6.12.117, 6.18.63 and 7.3.4.
Event log
- NoisyTCP (CVE-2026-99288) has existed in the Linux kernel since October 2015 and is rated 7.5/10 by Red Hat.
- Fixed in Linux stable kernels 6.12.117, 6.18.63 and 7.3.4.
- Default Nginx, Apache and SSH installs on QEMU/KVM VMs crashed within five seconds to two minutes in tests.
- A bare-metal Nginx server survived roughly 14 hours against a tuned proof-of-concept.
- Interim workarounds: disable SYN cookies per network namespace, or reduce to a single RX queue with RPS and RFS off.
By Tom Whitfield
A single crafted packet stream has been able to crash a default-configured, internet-facing Linux server since October 2015. That is the uncomfortable arithmetic behind NoisyTCP, a kernel flaw tracked as CVE-2026-99288 and rated 7.5 out of 10 by Red Hat, disclosed this week by security researcher Hyunwoo "V4bel" Kim.
The bug has shipped in virtually every kernel release for over a decade. It targets the TCP connection lookup path, where Kim found a race condition involving SYN cookies — the kernel's standard defence against connection-flooding attacks. An unauthenticated attacker who abuses that mechanism can bring down a remote machine.
The blast radius is not universal, but it is close to it for anyone running network services. Unpatched systems are exposed when they have multiple receive (RX) queues or Receive Packet Steering (RPS) enabled — both common, both spreading incoming packets across processor cores for throughput.
What does this change for virtualised estates?
Crash times vary wildly, and the pattern should concern anyone running workloads in VMs. Kim tested default installations of Nginx, Apache and SSH on QEMU/KVM virtual machines. They crashed within five seconds to two minutes.
A quiet bare-metal server running Nginx survived roughly 14 hours, even against a proof-of-concept tuned for that specific environment. The gap comes down to timing: hypervisors can pause a virtual processor mid-operation, widening the narrow race window the attack depends on. Density, in short, is a liability here.
Kim does not claim NoisyTCP enables remote code execution. He does note a second, harder-to-trigger path that causes a remote use-after-free memory error. "But remote DoS can be achieved without much difficulty," Kim said.
Which kernels fix it?
Linux stable kernels 6.12.117, 6.18.63 and 7.3.4 carry the fix. Patching is the recommended route, and the short version of this story is: schedule it now.
For administrators who cannot patch immediately, two workarounds exist, each with strings attached:
- Disable SYN cookies. This blocks the crash path but must be applied inside each container's network namespace individually, and it weakens protection against SYN flood attacks. It also does nothing against the second, harder-to-trigger memory corruption path.
- Reduce the receiving network interface to a single RX queue and switch off RPS and Receive Flow Steering (RFS). Kim said this configuration blocks the race outright. The trade-off is performance: you are deliberately collapsing the parallelism those features provide.
Neither workaround is free. Both trade availability engineering — flood resilience or packet distribution — for crash immunity while the patch waits in the change window.
Who found it?
Kim is not a newcomer to this territory. Earlier this year he reported "Dirty Frag", a Linux local privilege escalation flaw. NoisyTCP continues the pattern of long-dwelling kernel bugs surfacing years after introduction: in this case, a bug older than containerisation's mainstream adoption.
The practical checklist for infrastructure teams is short. Inventory internet-facing Linux hosts, check RX queue counts and RPS status, and get to 6.12.117, 6.18.63 or 7.3.4. If the patch queue is long, decide whether SYN-flood exposure or reduced network throughput is the cheaper temporary loss. Eleven years is a long time for a crash button to sit in the default build.
via itnews.com.au (Source)
More from Tom Whitfield
Show full bio
Market editor covering people, launches and day-to-day practice in enterprise information technology and the leaders who run at CIO Monitor.
133 articles
The newsroom digest
One email a day with the stories that matter. No noise.
Cross-references
- PwnEscape Flaw Escapes Firecracker MicroVM to Root on Host Kernel
- Apache Patches Critical HTTP/2 Remote Code Execution Flaw
- Check Point rushes hotfixes for exploited root privilege zero-day
- N-able N-central Zero-Day Exploited Within Hours; 21 Servers Mining Monero
- ConnectWise Suspends Cloud Services Following Cyberattack