Linux TCP Flaw Exposes MPTCP Listeners
CVE-2026-98432 is an IPv4 TCP use-after-free with a demonstrated root exploit against Linux servers using MPTCP listeners.
CVE-2026-98432 can give an unauthenticated remote attacker root command execution on Linux servers with a multipath TCP listener, according to an exploit Anthropic described on October 27. The Linux kernel CVE team assigned the TCP flaw a 9.8 CVSS score, while several distribution vendors rate it lower.
The vulnerability affects Ubuntu, Red Hat Enterprise Linux and other distributions. It was introduced in Linux 4.4 and sits in the IPv4 TCP connection setup path. IPv6 avoids the race because its implementation copies the relevant IP options.
The race occurs during the TCP handshake. After a connection completes, the kernel transfers IP options associated with the pending request to the new socket without copying them. If that socket is destroyed while another CPU is using those options to respond to a retransmitted connection request, the second CPU can read memory that has already been freed.
An attacker can induce both sides of the race by sending a TCP connection request containing IPv4 Record Route or Timestamp options and then resending the request. The kernel team says the resulting use-after-free can disclose nearby kernel memory, corrupt memory or crash the system.
The upstream fix makes the kernel wait for other CPUs to finish reading the options before freeing them. Kernel developers applied it on October 5, ten days before publication of the CVE.
Anthropic turned the race into root execution
Anthropic said its Claude model produced a two-stage exploit after initial prompting. The first stage repeatedly triggered the race until a kernel response disclosed a kernel address, defeating KASLR. In Anthropic’s test, obtaining that address took 48 minutes.
The second stage used the bug to overflow a buffer and replace a callback pointer with the address of a payload already placed in kernel memory. The kernel then invoked that pointer, executing a shell command chosen by the attacker as root. Anthropic said this stage took another five minutes.
A failed attempt only terminated that connection attempt rather than crashing the system, according to Anthropic, allowing repeated tries. The company did not publish exploit code or identify the tested kernel version and configuration.
Anthropic’s exploit targeted a server with an MPTCP listener. Neither the CVE record nor the kernel fix mentions MPTCP, and Anthropic did not say whether the same takeover works against servers without an MPTCP listener. Red Hat says a remote attacker able to reach an exposed TCP service can trigger the underlying race.
MPTCP allows a connection to use multiple network paths. Linux applications request it when opening a socket. Go programs have requested MPTCP by default for listening sockets since Go 1.24, according to the Go Security Team, making Go service inventories particularly relevant when assessing exposure.
Vendor scores reflect exploitability judgments
The kernel team treats attack complexity as low because the attacker induces both racing operations. Red Hat’s public CVE page rates the issue Moderate at 5.9, citing the difficulty of winning the race and describing a kernel crash as the most likely outcome. That page says confidentiality or integrity impact has not been demonstrated.
Red Hat’s machine-readable advisory reached a different conclusion on October 29: Important severity, an 8.1 score and high confidentiality and integrity impact. Its bug tracker also moved the issue from medium to high that day. Oracle and SUSE use the same 8.1 score, while Amazon assigns 7.0. The 8.1 vendor vectors differ from the kernel team’s assessment only by rating attack complexity high.
These differences should affect prioritization among vulnerable systems, not the decision to patch. Anthropic demonstrated that the race can be repeated and chained into root execution under at least one configuration.
Patch availability remains uneven
The upstream change has been backported to maintained stable kernel branches, but distribution users need the fix in their vendor kernel package. Debian, Oracle Linux, SUSE and Amazon Linux had released fixed packages for at least some kernels by October 29. Ubuntu and RHEL had not.
By October 30, Red Hat listed RHEL 8, 9 and 10 and OpenShift 4 as affected. Ubuntu also listed cloud, FIPS and real-time kernels as vulnerable or still being worked on. Neither vendor supplied a release date.
Operators should map exposed TCP services to their running kernel packages and distribution advisories, with explicit attention to MPTCP listeners and Go services. Install the vendor kernel update wherever one is available and complete the required reboot or workload replacement so the patched kernel is actually running.
For self-built kernels, fixed releases include 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3. The kernel CVE team recommends moving to the latest stable release rather than cherry-picking the individual change.
The Go Security Team documents GODEBUG=multipathtcp=0 as a runtime control and godebug multipathtcp=0 in the main module’s go.mod as a build-time control. Calls to SetMultipathTCP override both. It describes program-by-program disablement as fragile because one exposed MPTCP listener preserves the demonstrated attack path, and the kernel bug itself is not limited to MPTCP.
No vendor advisory identifies a generally deployable mitigation, and the sources do not establish that disabling MPTCP or filtering connection requests carrying IP options removes the broader risk. Systems awaiting vendor packages therefore need close exposure tracking, rapid update deployment and verification of the kernel running after maintenance.
Contains a referral link.
Keep Reading
KVMYour hypervisor was never a wall
Januscape (CVE-2026-53359) is a 16-year-old use-after-free in KVM/x86's shadow MMU that lets a guest panic or take over the host on Intel and AMD.
LXDLXD Flaw Enables Keyless Client Impersonation
CVE-2026-112816 let remote attackers impersonate trusted LXD and MicroCloud clients without possessing their private keys.
NorthflankNorthflank Breach Exposes AI API Keys
Northflank warned customers to rotate stored secrets after a breach led to reported unauthorized use of AI API keys.
Latest on the Wire
Full wire →- AI Investment Shows No Profit Impact Outside Tech SectorHacker News
- Engineers Build Open-Source Answers to Expensive SaaSHacker News
- LLM Bug Discovery Delays PostgreSQL 19 ReleaseHacker News
- LLM-Powered Infostealer VOLTA Automates Victim ValuationThe Hacker News
New signal daily · RSS
Stay in the loop
New writing delivered when it's ready. No schedule, no spam.