CVE-2026-98432: Real RHEL and OpenShift Risk Behind HAProxy
Assess the Linux TCP use-after-free flaw, Anthropic's MPTCP root exploit, Red Hat severity differences, and how full TCP proxies change exposure for OpenShift clusters.
Approximately 9 min read
CVE-2026-98432: A Critical Linux TCP Flaw, an AI-Built Root Exploit, and the Real Risk to OpenShift
Cybersecurity & AI | Technical Risk Briefing
Published 30 October 2026 | Evidence cutoff: 29 October 2026 (Eastern Time)
Scope: RHEL, Red Hat OpenShift, network load balancers, and enterprise vulnerability management
Executive assessment
CVE-2026-98432 is a real use-after-free vulnerability in the Linux kernel’s IPv4 TCP connection-establishment path. It affects kernel builds used by RHEL 8, 9 and 10 and Red Hat Enterprise Linux CoreOS (RHCOS) nodes in OpenShift Container Platform 4. The Linux kernel CVE authority assigned a CVSS v3.1 score of 9.8. On 27 October, Anthropic disclosed a laboratory exploit that achieved unauthenticated remote command execution as root against a Linux server with a Multipath TCP (MPTCP) listener.
For enterprise OpenShift deployments, the right question is not whether a worker node has a public IP address. In a conventional architecture it should not. The decisive questions are where the external TCP connection is terminated, which Linux kernel actually processes attacker-controlled TCP handshakes, and whether any vulnerable host remains directly reachable through internal or forwarded paths.
Assessment: Prioritize exposed Linux load balancers, MPTCP-enabled services and any packet-forwarded paths into vulnerable nodes for immediate review. Treat conventionally isolated OpenShift workers behind a genuine full TCP proxy as a different, lower-exposure scenario—not as automatically exploitable from the Internet, and not as immune to lateral movement. A 9.8 base score alone does not justify the same emergency change for every cluster.
What is confirmed—and what is not
| Question | Evidence-based answer |
|---|---|
| Is the CVE real? | Yes. Published by the Linux kernel CVE team on 15 October 2026. |
| Can it cross the network? | Yes, where an attacker can reach and influence the relevant IPv4 TCP handshake path. |
| Is the impact limited to denial of service? | No longer a safe general claim: Anthropic reports a working root RCE in a specific MPTCP-listener lab setup. |
| Does that prove RCE against every RHEL or OpenShift installation? | No. The published demonstration does not establish identical exploitability across vendor backports, kernel configurations and deployment topologies. |
| Is the issue patched upstream? | Yes. Upstream stable-branch fixes were published; downstream vendor patch status must be checked independently. |
| Is in-the-wild exploitation confirmed? | No confirmed campaign was established in the primary sources reviewed for this briefing. Absence of confirmation is not proof of absence. |
The underlying error is a race involving IPv4 TCP request-socket options. One CPU can continue using option memory while another path frees it, leaving the first CPU with a stale pointer. The upstream fix defers freeing the memory through an RCU grace period. This is a kernel-level flaw: updating an application container image does not repair its host kernel.
Anthropic says its exploit first leaked a kernel pointer to defeat KASLR, then achieved control-flow hijacking and root command execution. Its reported laboratory timing was approximately 48 minutes for the pointer leak plus five minutes for command execution. This is evidence of a feasible exploit chain under the tested conditions—not a measured compromise time for production OpenShift.
The load balancer is the boundary that matters
Consider a common enterprise request path:
Internet client
|
v
F5 / HAProxy / enterprise edge
|
v
OpenShift router / ingress endpoint
|
v
Service -> application pod
|
v
RHCOS host kernel
This drawing conceals a critical implementation distinction:
Full TCP proxy. HAProxy operating as a conventional reverse proxy, including its mode tcp, accepts the client connection and establishes a separate backend TCP connection. TCP payloads can be passed through without decrypting TLS, but the original client SYN, TCP options and IP options are not simply forwarded as the same handshake to the OpenShift node. This substantially constrains an Internet-origin attack that depends on manipulating those low-level handshake properties on the backend kernel. TLS passthrough does not mean TCP handshake passthrough.
Packet-forwarding load balancer. A device configured for destination NAT, direct server return (DSR), IPVS forwarding or similar packet-level delivery can allow the backend kernel to process the original connection handshake. Its impact cannot be assessed using the full-proxy assumption. The VIP’s actual implementation matters more than whether the equipment is marketed as “Layer 4.”
The first-hop Linux host still matters. If an Internet-facing HAProxy instance itself runs on an affected RHEL/Linux kernel, the external client handshake is processed by that proxy host’s kernel. Protecting OpenShift backends does not automatically protect the Linux load balancer. Appliances running other operating systems require their own product-specific assessment.
Internal reachability remains relevant. A compromised workload, a host on a permissive internal segment, or an exposed NodePort/hostNetwork listener may be able to reach the vulnerable node directly. OpenShift ingress routers may use host networking or node ports depending on the publishing strategy. Worker nodes need not be Internet-facing for a network-reachable kernel flaw to matter.
Importantly, none of these observations demonstrates that ordinary HTTP traffic relayed through a full TCP proxy can reproduce Anthropic’s MPTCP-specific exploit. That remains an evidence gap, not an assumption to fill with worst-case speculation.
How to rank enterprise exposure
| Scenario | Suggested response | Rationale |
|---|---|---|
| Internet-exposed RHEL/Linux TCP proxy with an affected kernel | Urgent validation | The proxy host, not just its backends, handles untrusted TCP handshakes. |
| Directly reachable MPTCP listener on an affected Linux host | Urgent validation and containment | Closest to the conditions of the publicly described root-RCE PoC. |
| OpenShift ingress reached through packet-level NAT/DSR forwarding | High-priority architecture review | Original connection-handshake semantics may reach a vulnerable RHCOS kernel. |
| OpenShift behind a true full TCP proxy; node ports restricted to trusted networks | Accelerated patch planning and internal exposure review | External exploit path is constrained, but internal paths and vendor patching remain relevant. |
| Segmented, non-production RHEL servers without relevant reachable services | Routine accelerated remediation | Lower demonstrated exposure, subject to inventory findings. |
These are operational priorities, not CVSS scores. They assume the affected kernel is present and must be revised if exploit conditions, reachability or vendor guidance change.
Red Hat severity data deserves a second look
The Linux kernel CVE authority scores the vulnerability 9.8. Red Hat’s CVE page has returned different scores across retrieved versions: 5.9 with denial-of-service-only impact in one published response, and 8.1 with high confidentiality, integrity and availability impact in another. Both rate attack complexity as high. As of this briefing’s evidence cutoff, Red Hat’s narrative continued to characterize the practical risk primarily as denial of service. Because the published scoring surfaces have not consistently agreed, treat the exact vendor CVSS snapshot as time-sensitive and confirm it in the current Red Hat advisory before escalating operational priority. Do not silently promote either score to an environment-specific risk rating.
The 27 October Anthropic disclosure is especially relevant to that discrepancy. It challenges the categorical proposition that this CVE has no demonstrated confidentiality or integrity impact, while not proving the same exploit works on every RHEL/RHCOS build. Security teams should request product-specific clarification from Red Hat rather than treating either 5.9 or 9.8 as an environment-specific risk rating.
At this cutoff, Red Hat’s published information identified RHEL 8/9/10 and OpenShift 4 RHCOS nodes as affected, with fixes described as deferred in the product-security statement retrieved for this briefing. Check current RHSA advisories and the exact product release before reporting any system as unpatched. RHEL 7 is identified by Red Hat as not affected. The upstream patch does not mean a supported RHCOS update is already available.
Recommended actions: next 48 hours
- Map the true TCP boundary. For each external VIP, confirm full TCP termination versus NAT/DSR/packet forwarding. Include infrastructure routers, management endpoints, NodePort services and the operating system of Linux-based load balancers.
- Inventory affected kernels, not merely container images. Capture OpenShift node kernel versions and RHCOS release details, along with RHEL kernel packages on front-end proxies, bastions and other network-facing hosts.
- Identify MPTCP and alternative handshake exposure. Check whether MPTCP listeners actually exist and whether untrusted networks can reach them. Review
net.ipv4.tcp_migrate_reqand relevant reuseport/BPF usage where applicable. An enabled MPTCP sysctl alone does not prove a listener exists; disabling MPTCP alone is not a vendor-verified fix for the entire CVE. - Review east-west paths. Confirm that cluster nodes, host-network listeners and management ports are not broadly reachable from application networks or untrusted zones. Validate network policy and infrastructure ACLs against actual routing, not architecture diagrams alone.
- Escalate vendor clarification. Ask Red Hat for affected kernel build numbers, whether the Anthropic PoC changes its RHEL/RHCOS assessment, supported mitigations and expected RHSA/OpenShift update availability.
- Prepare a supported rolling update. For OpenShift, use the vendor-supported cluster/RHCOS update process with capacity and disruption checks; do not manually replace the host kernel or cherry-pick an upstream patch into production RHCOS.
Useful read-only inventory commands:
oc get nodes -o custom-columns=NAME:.metadata.name,OS:.status.nodeInfo.osImage,KERNEL:.status.nodeInfo.kernelVersion
oc get ingresscontroller -n openshift-ingress-operator -o wide
oc get pods -n openshift-ingress -o wide
On an authorized Linux host, inspect the relevant settings and listeners:
uname -r
sysctl net.mptcp.enabled net.ipv4.tcp_migrate_req
ss -lntp
ss -Mlntp
The MPTCP-specific commands may be unsupported on older kernel/tool combinations; missing support or an empty MPTCP listing does not establish that the underlying kernel CVE is absent. These commands are inventory aids, not an exploitability test.
The AI-security significance
The most consequential part of this story is not the numerical CVSS disagreement. Anthropic publicly described an AI-assisted exploit chain that moved from a kernel memory-safety race to KASLR defeat and root code execution in a bounded laboratory test. That narrows the credibility gap between “theoretical kernel memory corruption” and a working, multi-stage remote exploit.
Two conclusions follow for defenders. First, exploit-development capability can change faster than enterprise patch and exception cycles. Second, threat intelligence must capture new exploit evidence and affected network paths, not just new CVE identifiers. An enterprise may reasonably defer an emergency OpenShift restart behind a verified full TCP proxy while still treating a vulnerable Linux edge proxy or reachable MPTCP listener as urgent. Those are not contradictory decisions; they reflect different attack surfaces.
Bottom line: CVE-2026-98432 warrants serious attention across RHEL and OpenShift, but it is not evidence that an Internet user can automatically compromise ordinary OpenShift workers behind a properly implemented full TCP proxy. Verify the first TCP termination point, the kernel running there, direct and lateral reachability, and product-specific vendor guidance before choosing the change window.
Primary sources
- Linux kernel CVE announcement, 15 October 2026 — https://lists.openwall.net/linux-cve-announce/2026/10/15/1
- Red Hat Product Security, CVE-2026-98432 — https://access.redhat.com/security/cve/cve-2026-98432
- Anthropic Frontier Red Team, Discovering and exploiting vulnerabilities in critical, real-world software with Claude, 27 October 2026 — https://www.anthropic.com/research/discovering-vulnerabilities-with-claude
- Ubuntu CVE tracker, CVE-2026-98432 — https://ubuntu.com/security/CVE-2026-98432
- HAProxy documentation, TCP proxy — https://cdn.haproxy.com/documentation/haproxy-configuration-tutorials/protocol-support/tcp/
- HAProxy ALOHA documentation, proxy versus LVS packet forwarding — https://cdn.haproxy.com/documentation/haproxy-aloha/load-balancing/overview/
- Red Hat OpenShift documentation, ingress publishing strategies — https://docs.redhat.com/en/documentation/openshift_container_platform/4.21/html/ingress_and_load_balancing/configuring-ingress-cluster-traffic
- Red Hat OpenShift documentation, RHCOS update model — https://docs.redhat.com/en/documentation/openshift_container_platform/4.16/html/architecture/architecture-rhcos
Editorial note: This is an evidence-based technical assessment, not a penetration-test finding or an assertion of exploitation in any specific enterprise environment. Product and patch statuses should be rechecked before operational use.