Critical Linux Kernel TCP Flaw CVE-2026-98432 Still Unpatched on Ubuntu and RHEL

Ubuntu and Red Hat Enterprise Linux have yet to ship fixed kernels for CVE-2026-98432, a critical Linux kernel TCP flaw that Anthropic says its unreleased AI model turned into a zero-click remote root exploit.

The bug is a use-after-free in the kernel’s IPv4 TCP handshake code, and the Linux kernel’s own CVE team rates it 9.8 out of 10 on the CVSS scale. It has been present since Linux 4.4, according to the CVE record, so it reaches nearly every maintained distribution kernel. Red Hat says RHEL 7 is the exception because its kernel never received the vulnerable code.

Anthropic said on Tuesday that its internal Claude research model wrote an exploit that lets a remote attacker with no credentials run commands as root on Linux servers with a multipath TCP listener. Our earlier report covers the Chrome, Safari, Android, and Windows chains from the same research.

As of Thursday, Debian, Oracle Linux, and some SUSE and Amazon Linux kernels are fixed. Ubuntu lists every supported release as vulnerable, and Red Hat has deferred its fix to a future update.

How the TCP Race Condition Works

When a Linux server completes a TCP handshake, the function tcp_v4_syn_recv_sock() moves the IP options stored with the pending connection request to the newly created socket. It hands over the original rather than a copy, so two objects now point to the same memory.

The kernel processes those pending requests without holding the listener lock. A second CPU core can therefore be reading the same options to answer a retransmitted SYN while the first core destroys the new socket and frees it, and the second core ends up reading freed memory.

The upstream fix, written by Eric Dumazet and crediting Anthropic researcher Xinyang Ge as the reporter, makes the kernel wait until all readers have finished before freeing the options. The same patch closes a second path to the bug through TCP request socket migration, which is enabled by the net.ipv4.tcp_migrate_req setting or a BPF reuseport program. IPv6 is not affected by the main race because its function version copies the options.

How Anthropic Turned the Bug Into a Root Exploit

Anthropic’s exploit works in two stages. It first wins the race repeatedly, so the kernel reads an attacker-controlled length field and sends extra kernel memory back in its replies. Once a reply contains a kernel pointer, the attacker can defeat KASLR, the defence that randomises where the kernel is loaded in memory.

The same bug is then used to overflow a buffer and overwrite a callback pointer with the address of a payload already staged in kernel memory. When the kernel calls it, the payload runs a shell command of the attacker’s choice as root.

A lost race fails only that one connection attempt and does not crash the server, so an attacker can keep retrying. In Anthropic’s test, leaking a kernel address took 48 minutes, and reaching root command execution took another five.

Neither the CVE record nor the kernel fix mentions multipath TCP (MPTCP), and Anthropic has not said why its exploit targets MPTCP listeners or whether servers without one can be taken over the same way. It has not released the exploit code.

Why Vendors Disagree on the Severity

The kernel CVE team updated its record on October 24 with a score of 9.8. Its scoring scenario assumes an unauthenticated attacker who sends a TCP SYN carrying IPv4 Record Route or Timestamp options and then resends it to hit the race, which the team treats as low attack complexity.

Red Hat takes the opposite view. Its CVE page rates the flaw as Moderate, with a score of 5.9, citing the difficulty of winning the race, and notes that the residual risk is limited, with “denial of service being the practical worst case.” That judgement sits awkwardly beside Anthropic’s working root shell.

Red Hat’s machine-readable VEX data was updated on Thursday to rate the flaw as Important with a score of 8.1, while the public CVE page still showed it as Moderate. Cyber Kendra could not independently confirm the VEX change.

Other vendors sit in between. SUSE and Oracle score it 8.1 and rate it Important, Amazon scores it 7.0, and Ubuntu uses the kernel team’s 9.8 with a High priority.

Which Linux Distributions Have a Fix

Upstream, the fix is committed at e3f33b0a1d89 and ships in stable kernels 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, and 7.2.10, and in mainline 7.3 release candidates from 7.3-rc7 onward. Ubuntu’s tracker shows fixed builds in preparation for its 22.04, 24.04, 26.04 and 26.10 kernels, along with its cloud, FIPS and real-time variants, but gives no release date. Red Hat has given none either.

Is CVE-2026-98432 Being Exploited?

No vendor advisory reports attacks using the flaw, and it is not in CISA’s Known Exploited Vulnerabilities catalogue. Anthropic’s exploit has not been published.

Public research is starting to appear. On Thursday, a security research group called Innora published a repository containing a root cause analysis and a QEMU-based reproduction environment for CVE-2026-98432, marking the full chain as verified in its lab. The group says it will not release a weaponised exploit in the short term. Cyber Kendra has not tested the material.

What Linux Administrators Should Do Now

Where a fixed package exists, install it and reboot, because a kernel update takes effect only when the new kernel is running. Running uname -r shows the active kernel version, which can be compared against the fixed releases above or the vendor’s advisory.

Teams that build their own kernels should move to one of the fixed stable releases. The kernel CVE team advises updating to the latest stable kernel rather than cherry-picking individual fixes.

Ubuntu and RHEL users do not yet have a fixed kernel to install. Red Hat says no available mitigation meets its standards for ease of deployment and stability, and no other vendor advisory lists one. Neither the advisories nor Anthropic’s post says whether disabling multipath TCP would reduce the risk.