The service is a doorway to the kernel

Security researcher Hyunwoo Kim disclosed NoisyTCP, tracked as CVE-2026-99288, after a coordinated embargo ended. It is a race in the Linux kernel's TCP socket lookup path. In the demonstrated case, a stale lookup can return a newly reused request socket with no listener attached. The resulting null-pointer dereference panics the kernel.

The attacker does not need an account or a specially configured application. Kim says a reachable, ordinary TCP listener such as a web server or SSH can be enough when the host has multiple receive queues or packet steering across CPUs. In his tests, crashing a virtual machine took seconds to minutes; a quiet bare-metal server took much longer. Those timings are laboratory results, not a promise about every deployment.

This is not a confirmed remote-code exploit

Kim demonstrated remote denial of service and says a general-purpose remote-code exploit has not been established. Do not fold this CVE into the separate Linux TCP issue, CVE-2026-98432, or describe NoisyTCP as an observed intrusion campaign.

Find the running kernel, then follow your vendor

  1. 1

    Map reachable hosts. Identify Linux servers and container hosts with externally reachable TCP listeners. Record the distribution, release, installed kernel package and running kernel.

  2. 2

    Check fixed packages. Use your distribution's advisory rather than an upstream version alone: vendors may backport the fix. Debian's tracker lists fixed security packages for Bookworm and Trixie. Plan the kernel update and required reboot.

  3. 3

    Verify after restart. Confirm that the running kernel contains the vendor fix. Keep hosts without a released fix in an exception list and limit unnecessary network reachability while monitoring vendor guidance.

A workaround has a cost

Kim describes disabling TCP SYN cookies or reducing packet processing to one receive queue as temporary ways to obstruct the race. Disabling SYN cookies weakens protection against SYN floods, and queue changes can affect performance. Those controls need operational review; a supported, fixed kernel is the durable answer. A container does not isolate the impact here: a kernel panic can take down the host and its other workloads.