CVE-2026-98432: Linux Kernel Exploit Puts Go Servers at Risk
Anthropic demonstrated a remote root exploit for Linux kernel CVE-2026-98432. Learn why Go 1.24 MPTCP defaults matter, which kernels are fixed and what administrators should check.
On this page
A Linux kernel flaw disclosed in October has moved from a memory-management bug to a demonstrated remote root exploit. Anthropic says an unreleased research model built an attack for CVE-2026-98432 that can compromise a Linux server running a Multipath TCP listener without credentials or user interaction. The disclosure matters beyond specialist networking setups: Go has enabled Multipath TCP on Linux listening sockets by default since version 1.24. Here is how the flaw works, which systems need attention, and why disabling one networking feature is not a substitute for a kernel update.
What changed when Anthropic demonstrated the exploit
The vulnerability itself was disclosed by the Linux kernel security team on October 15, 2026. The more consequential development came on October 27, when Anthropic published details of an exploit produced with an unreleased internal research model. The company described a two-stage attack against Linux servers with a Multipath TCP listener: first, the exploit leaks kernel memory to defeat address randomisation; then it uses the same underlying flaw to take control of kernel execution and run a command as root. Root access gives an attacker the highest level of control over the operating system, potentially including access to files, processes and other services. These are Anthropic's reported proof-of-concept results, not evidence that attackers are already using the exploit in real-world intrusions.
Anthropic reported that its demonstration took 48 minutes to obtain a kernel pointer and another five minutes to reach command execution. Those timings describe its own test, not a guaranteed attack duration on every server. The company has not published the complete exploit, and the public disclosure does not establish that every vulnerable Linux machine can be compromised in the same way. The distinction is important: the research demonstrates a serious route to remote compromise, but it does not prove widespread exploitation.
How CVE-2026-98432 creates a use-after-free
The defect sits in the Linux kernel's handling of IPv4 Transmission Control Protocol (TCP) connections. TCP is the networking protocol commonly used by services such as websites, secure shell access and application programming interfaces. During a connection handshake, the kernel can store extra information from an incoming packet in memory associated with a pending connection. In a particular timing race, one CPU can still be reading that information while another has already freed it. The result is a use-after-free: software accesses memory after it has been released, potentially reading stale data or corrupting memory that has since been reused.
The kernel developers traced the problem to how IPv4 connection options are shared between a pending request and the new connection created from it. A retransmitted connection request can be processed at the same time as the original handshake is being completed. If the new connection is destroyed while another CPU is still reading the shared options, the memory can be freed too early. A similar race can occur when pending connection requests are migrated between sockets under certain configurations. The upstream correction changes the freeing operation to use kfree_rcu(), which delays releasing the memory until readers using the kernel's Read-Copy-Update mechanism have finished.
Why Go 1.24 makes the MPTCP detail relevant
Anthropic's published root exploit targeted a server with a Multipath TCP (MPTCP) listener. MPTCP extends ordinary TCP so a connection can use multiple network paths, rather than relying on a single path for its entire lifetime. It is useful in environments where devices have more than one network connection, but administrators may not realise that an application is using it. Go's official release notes state that the net package began using MPTCP by default for listening sockets on supported systems in Go 1.24. That means a Go service can expose the demonstrated attack path without its developers explicitly setting out to enable MPTCP.
The Go Security Team addressed this connection directly in an advisory circulated on October 29. It confirmed that Go programs enable MPTCP by default for Linux listening sockets from Go 1.24, and explained that developers can change the default with the multipathtcp GODEBUG setting. For example, setting GODEBUG=multipathtcp=0 disables the default at runtime, while adding godebug multipathtcp=0 to the main module's go.mod file changes the build-time default. The team also warned that turning off MPTCP in individual programs is a fragile response: one overlooked listener could leave the server exposed, and the underlying kernel bug is not limited to MPTCP.
The flaw is broader than the published MPTCP exploit
It would be a mistake to interpret Anthropic's demonstration as proof that only MPTCP servers are affected. The Linux kernel advisory describes a flaw in the ordinary IPv4 TCP connection-handling code, with request-socket migration providing another route to the same memory-management problem. The Go Security Team explicitly cautioned that systems without MPTCP listeners could still be vulnerable to similar attacks. Disabling MPTCP may remove the specific condition used in Anthropic's demonstration, but it does not repair the vulnerable kernel code. Administrators should therefore treat MPTCP checks as an exposure assessment, not as a declaration that a server is safe.
The distinction also matters when reviewing application inventories. A server may run a Go-based API, proxy or internal service whose listener configuration was inherited from the language runtime rather than chosen by an administrator. Conversely, simply finding Go installed does not prove that a vulnerable MPTCP listener exists: the application must use a relevant listener, and the operating system must support the feature. The useful question is not just whether a machine runs Go, but which processes are listening, how those sockets are configured and whether the running kernel contains the correction.
Which Linux kernel versions contain the upstream fix
The Linux kernel security announcement says the issue was introduced in version 4.4 and lists corrections for several maintained stable branches. The fixed upstream versions include 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56 and 7.2.10, with the correction also included in the 7.3 release-candidate line. These numbers are useful for administrators building or maintaining kernels directly. They are not a universal checklist for distribution packages, because Linux vendors routinely backport security fixes into kernels that retain different version numbers.
Distribution status was uneven at the end of October. Ubuntu's security tracker, updated October 30, still listed the fix as work in progress for several supported releases, including Ubuntu 22.04, 24.04, 26.04 and 26.10. Red Hat lists affected Enterprise Linux 8, 9 and 10 products, as well as relevant OpenShift systems, and says that no mitigation currently meets its security team's criteria. Other vendors have shipped fixes for at least some products: Oracle Linux published kernel errata on October 16, while Amazon Linux 2023 marked its optional 6.12 kernel fixed on October 29 and its 6.18 kernel fixed earlier. The status must be checked against the exact distribution and kernel package running on each machine.
What server administrators should check now
Start by identifying the kernel actually running on each Linux host, rather than relying on the version listed in a deployment image or package inventory. The command uname -r prints the active kernel release. Next, check the security tracker for the distribution and release installed on that machine, searching for CVE-2026-98432. This matters particularly on Ubuntu and Red Hat systems, where the upstream correction may exist while the vendor package is still pending. If the vendor has released a fixed package, install it through the supported update channel and schedule the required reboot so the server actually starts using the corrected kernel.
Teams operating Go services should also review whether their applications use MPTCP listeners. The Go Security Team documents the multipathtcp setting and per-socket controls through net.ListenConfig.SetMultipathTCP, but those controls should not become the main remediation plan. They can help identify or reduce exposure to the demonstrated MPTCP route while a distribution prepares its kernel update. They cannot rule out other ways to trigger the underlying IPv4 TCP flaw. Red Hat's advisory specifically says that it has no mitigation meeting its deployment and stability requirements, reinforcing the need to obtain the vendor's corrected kernel rather than rely on an improvised workaround.
What the disclosure does not establish
There are limits to what can responsibly be concluded from the public evidence. Anthropic has described a working exploit in its own research environment, but has not released the exploit code or shown that the same chain works against every affected distribution and configuration. The published research does not establish that attackers are exploiting CVE-2026-98432 in the wild. Nor does the fact that the demonstrated attack used MPTCP mean that ordinary IPv4 TCP servers are unaffected. Administrators should separate the confirmed kernel defect, Anthropic's specific exploit demonstration and the still-uncertain scope of practical exploitation.
The next meaningful change will be the arrival of corrected packages for distributions that still list the vulnerability as unresolved. Until then, server owners should keep their exposure inventory current, watch their vendor's CVE status and avoid treating a disabled MPTCP setting as a permanent fix. For Go teams, the runtime default is now part of the review: check the listener behavior of deployed binaries, especially services built with Go 1.24 or later, while the kernel update remains the priority.
Share this post



