TCP Use-After-Free: CVE-2026-98432 / CVSS 9.8

#Overview

Linux's stable kernels shipped a fix on October 11 and 15, 2026 for CVE-2026-98432, a use-after-free in the IPv4 TCP handshake that a remote, unauthenticated equal can reach. The kernel's own CVE Numbering Authority (CNA) scores it CVSS 3.1 9.8, Critical. The fixed releases are 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10, and 7.3, and the bug goes rear to Linux 4.4. The upstream fix is a four-line change, but most servers run a distribution's kernel, and as of 8:00 PM CDT, October 29, Ubuntu's tracker motionless says "Vulnerable" for its kernel on all release, and Red Hat lists RHEL 8, 9, and 10 as affected alongside no fix.

Why it's urgent: the bug sits in code all Linux server runs whenever it accepts an IPv4 TCP connection, before any use sees the connection. Anthropic, whose investigator reported it, says its Claude example wrote a operating proof-of-concept utilize that runs commands as base on "Linux servers alongside a multipath TCP listener," alongside no credentials. See the timeline.

#Key facts

#Who is affected

Any Linux kernel from 4.4 on that lacks the fix, whether it serves the network immediately or hosts containers: a receptacle runs on the host's kernel, so a new depiction doesn't alter anything. IPv6 connections aren't affected by the chief issue; the commit says "IPv6 is not affected by the archetypal issue."

#Distribution status

Distribution kernels transport backported fixes under their own type numbers, so inspect the tracker, not the upstream number. As of 8:00 PM CDT, October 29:

• Ubuntu: the tracker lists the linux bundle as "Vulnerable" on all release. On 26.10, 26.04, 24.04, and 22.04 it adds "work in progress" in 7.3.0-10.10, 7.0.0-43.43, 6.8.0-149.149, and 5.15.0-203.213; 20.04, 18.04, and 16.04 display lone "Vulnerable." Many haze and hardware-enablement kernels are listed the identical way, and no Ubuntu Security Notice has been published.
• Red Hat: RHEL 8, 9, and 10 and OpenShift Container Platform 4 are listed as affected, alongside no fix; RHEL 6 and 7 aren't affected (CVE page). Red Hat says mitigation "is either not accessible or the currently accessible options do not encounter the Red Hat Product Security criteria."
• Debian: fixed in 12 "bookworm" (6.1.190-1, DLA-4841-1, and linux-6.12 6.12.113-1~deb12u1, DLA-4850-1), 13 "trixie" (6.12.113-1, DSA-6568-1), and unstable (tracker).
• SUSE: rated important. SUSE Linux Enterprise Server 16.0 (kernel-default 6.12.0-160000.40.1) and the 15 SP4, SP5, and SP6 LTSS kernels are fixed; 15 SP7 is in progress, and openSUSE Leap 15.6 and 16.0 are listed as affected (CVE page).
• Oracle Linux: rated Important, CVSS 3.1 8.1. Unbreakable Enterprise Kernel (UEK) errata for Oracle Linux 7, 8, 9, and 10 came out October 16: ELSA-2026-500397, ELSA-2026-500398, and ELSA-2026-500399 (CVE page). The leaf lists lone the UEK kernels.
• Amazon Linux: rated Important, CVSS 3.1 8.1. On Amazon Linux 2023, kernel6.18 is fixed (ALAS2023-2026-3251, October 24) and kernel6.12 is fixed (ALAS2023-2026-3335, October 29). The default kernel bundle and all Amazon Linux 2 kernel are "Pending Fix" (CVE page).

#How the exposure works

When a Linux server accepts a TCP connection, the kernel tracks the half-open handshake in a lightweight "request socket," afterward creates a complete socket formerly the handshake completes. If the connecting peer's IPv4 header carried options, the kernel keeps them to echo in its replies. The bug: the kernel handed those saved options from the petition socket to the new association without copying them, afterward freed them immediately whenever the association went away, during another CPU could motionless be study them to build a handshake reply. Reading freed kernel recollection is a use-after-free.

The equal needs no login, and it all happens inner the kernel's handshake handling, before any use code runs. The CNA's scoring notes say the old data can leak kernel heap recollection rear to the peer, steer kernel writes out of bounds, or collision the machine. Anthropic says a misplaced competition "only fails the handshake without crashing the entire system, so a resolute attacker can retry indefinitely until they succeed."

#The fix

Eric Dumazet's commit, "ipv4: liberated inet_opt and ireq_opt following an RCU grace period," changes two plain kfree() calls to kfree_rcu(), which waits until all audience has completed alongside the recollection before freeing it. Other code paths that substitute these options already did that. The commit additionally fixes the identical issue in request-socket immigration (net.ipv4.tcp_migrate_req=1, or a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE program), and says a connected IPv6 case "will be addressed in a distinct patch."

net/ipv4/af_inet.c

+1 −11 lines added, 1 removed

net/ipv4/tcp_ipv4.c

+1 −11 lines added, 1 removed

2 records changed+2 −22 lines added, 2 removed

#Which servers are exposed

• Every Linux presenter that accepts IPv4 TCP connections on an affected kernel: web servers, SSH, databases, burden balancers, and Kubernetes nodes. The CNA's notes say the susceptible way runs "during the TCP handshake, before use authentication."
• Hosts alongside a multipath TCP (MPTCP) listener first. That's the setup Anthropic's operating utilize needs: "Linux servers alongside a multipath TCP listener." Anthropic hasn't stated why that listener matters, and the CVE document doesn't citation MPTCP. A server has one lone if an use opens an MPTCP socket; the inspect is below.

Red Hat rates the bug lesser than the CNA. It scores the assault complexity as High (CVSS 8.1) and describes the outcome as chiefly a collision "or possible recollection corruption." Anthropic's lab utilize shows a distant base casing on at smallest one configuration, so a lesser ranking isn't a logic to wait.

#How to inspect your systems

#Check the kernel

Compare the outcome alongside your distribution's position above, not alongside the upstream version. If the installed kernel is newer than the operating one, the fix isn't energetic until you reboot. Run these on receptacle hosts and Kubernetes nodes, not inner containers: a receptacle reports the host's kernel. We ran them on a Debian 12 presenter and in a Rocky Linux 9 container.

#Check for multipath TCP listeners

ss -M lists lone MPTCP sockets, so an bare outcome method no MPTCP listener. net.mptcp.enabled = 1 solitary doesn't create one: an use has to open an MPTCP socket. We tested the two alongside iproute2 6.1 in a Debian 12 container, anywhere ss -Mlnp listed a test MPTCP listener and skipped a plain TCP one beside it.

#How to fix

#Update the kernel and reboot

Install your distribution's fixed kernel as shortly as it ships, afterward reboot: a kernel update doesn't obtain consequence until the device boots into it. Reboot receptacle hosts and Kubernetes nodes too. Appliances and haze images that bundle a kernel need their vendor's update.

On Debian, SUSE Linux Enterprise Server 16.0 and the 15 LTSS releases, Oracle Linux's UEK kernels, and Amazon Linux 2023's kernel6.12 and kernel6.18, the fix is out. On Ubuntu, RHEL, Amazon Linux 2, and Amazon Linux 2023's default kernel, it isn't yet, as of 8:00 PM CDT, October 29. Watch the trackers connected above. Running a self-built kernel? Move to the fixed stable publish for its line.

#If you can't place yet

Red Hat publishes no mitigation. The following is Atlas's interim recommendation, according to the CNA's scoring notes, which say the attacking equal supplies IPv4 Record Route or Timestamp options in its SYN. TCP traffic nearly never carries IPv4 options legitimately, so dropping TCP packets that have them should disbursal nothing:

ip hdrlength gt 5 matches any IPv4 header longer than the 20-byte minimum, which lone options produce. The regulation runs before association tracking and doesn't contact IPv6 or non-TCP traffic. We tested it in a Debian 12 container: it dropped a SYN carrying a Record Route option, and average TCP connections motionless worked. We haven't tested it against an exploit, and it doesn't fix the kernel. It lasts until reboot, which is fine if the reboot is into a fixed kernel; eliminate it alongside nft delete array ip cve_2026_98432. If your border firewall can autumn IPv4 packets alongside options, doing it there protects all presenter rearward it.

Until then, keep MPTCP listeners off the internet, and put the hosts that have one at the top of the place list.

#Timeline and exploitation status

All times are CDT (UTC−5).

Exploitation status. Anthropic says Claude wrote the utilize "after first prompting," that it runs "arbitrary commands as root, without any credentials or person interaction" on servers alongside a multipath TCP listener, and that in its evidence of idea "it takes 48 minutes to leak a kernel pointer and another five minutes to attain the casing command execution." Anthropic hasn't published the exploit. Its disclosure policy gives developers 90 days to place and a additional 45 days before Anthropic discusses particulars beyond a community advisory. By 8:00 PM CDT on October 29, eight GitHub repositories referenced the CVE. Most depict themselves as audit or mitigation tooling, among them an Ansible function and a livepatch for AlmaLinux and Rocky Linux. One, created at 7:22 PM, says it includes a competition trigger. The CVE isn't in CISA's Known Exploited Vulnerabilities catalog as of 8:00 PM CDT. We haven't tested any community evidence of concept, and we aren't linking them.

#Updates to this post

We'll revise this article and grade it updated whenever any of these change: fixed kernels from Ubuntu, Red Hat, or Amazon Linux; NVD's analysis; a KEV listing or any study of exploitation; the distinct IPv6 fix.

• October 29, 8:00 PM CDT: Added Oracle Linux, fixed in its UEK kernels since October 16, and Amazon Linux, fixed in Amazon Linux 2023's 6.12 and 6.18 kernels and pending elsewhere. Eight GitHub repositories now citation the CVE, one describing a competition trigger. Ubuntu, Red Hat, Debian, SUSE, NVD, and KEV are unchanged since 11:00 AM, and nothing was corrected.

#Get help

If the difficult part is knowing which hosts, nodes, and appliances run which kernel, and which of them can reboot this week, our advisors can assistance scheme it.

Atlas Service

Atlas Technical Consulting

Senior advisors for safety audits, code reviews, and DevOps activity — a two-hour finding obstacle alongside no retainer commitment, and a 24-hour reply SLA.

The bug is reachable through any listening TCP port, so it helps to cognize which ones the net can see. Our liberated external audit checks the domains and community IPs you verify for exposed services, publically apparent application versions alongside known CVEs, TLS, email security, DNS, and HTTP headers.

Atlas Service

Free External Security Audit

Non-destructive checks of your verified domains and community IPs, and a PDF study alongside all finding classified and a fix for each, inside 3 endeavor days of verification.

It's the second Linux kernel fix this duration that lone takes consequence following a reboot. The other, Fragnesia, is being chained alongside PHP2Shell to get root.

#Sources