As reported by The Hacker News, a critical use-after-free vulnerability in the Linux kernel's TCP connection setup code—CVE-2026-98432—has surfaced with a striking CVSS 9.8 rating from the kernel's own CVE team. What makes this particularly notable is that Anthropic has demonstrated an AI-generated exploit achieving unauthenticated remote root execution against servers running multipath TCP (MPTCP) listeners. This is not a theoretical risk; the exploit path is proven.
Vulnerability Summary
| Field | Detail |
|---|---|
| CVE ID | CVE-2026-98432 |
| CVSS (Kernel Team) | 9.8 — Critical |
| CVSS (Vendor Range) | 5.9 – 8.1 (Red Hat, Amazon, Oracle, SUSE) |
| Bug Type | Race condition / Use-after-free in IPv4 TCP connection setup |
| Exploitation Requirement | MPTCP listener enabled; remote, no credentials needed |
| Impact | Remote code execution as root |
| Active Exploitation | Not yet observed in the wild (exploit PoC exists from Anthropic) |
| Upstream Patch | Patched and backported to maintained stable branches |
| Reporter | Xinyang Ge (Anthropic) |
The Scoring Discrepancy Matters
The gap between the kernel team's 9.8 and Red Hat's 5.9 is not merely academic. It reflects a fundamental disagreement on exploitability assumptions—specifically whether MPTCP is common enough in production to warrant the higher score. Red Hat's lower rating implicitly assumes MPTCP deployment is niche. Shield53 disagrees with that assessment for a growing subset of infrastructure. MPTCP adoption has been quietly increasing in cloud environments, SD-WAN configurations, and mobile carrier networks where connection resilience across multiple paths is valued.
The threat model here is straightforward: any internet-exposed Linux server with an MPTCP-capable socket listener can be fully compromised by an unauthenticated attacker. That is a Critical-class exposure regardless of how common the configuration is.
Who Is Most Exposed
Patch Status by Distribution (as of Oct 29)
| Distribution | Status |
|---|---|
| Debian (trixie, bookworm) | Fixed — packages available |
| Oracle Linux (UEK) | Fixed — ELSA advisories released |
| SUSE / SLES 15 SP4-SP6 | Fixed — kernel 6.12.0-160000.40.1+ |
| Amazon Linux 2023 (kernel 6.18, 6.12) | Fixed — ALAS advisories released |
| Ubuntu (all supported LTS) | Work in progress — no package released |
| RHEL 8/9/10, OpenShift 4 | Fix deferred — no package released |
Shield53 Recommendations
Immediate Actions
- Audit for MPTCP exposure: Inventory all Linux servers and identify which have MPTCP enabled. Check with
sysctl net.mptcp.enabledand review socket configurations in running services. - Disable MPTCP where not required: On unpatched systems, set
net.mptcp.enabled=0via sysctl as an interim mitigation. This eliminates the known exploit path but validate application impact first. - Network-level controls: Restrict inbound TCP access to MPTCP-capable services using firewalls or security groups. The exploit requires network reachability to a listening socket.
- Prioritize patching for fixed distributions: If running Debian, Oracle, SUSE, or Amazon Linux, apply kernel updates immediately and reboot during the next maintenance window.
- Plan for Ubuntu/RHEL: Both vendors are working on fixes but have not shipped. Monitor their security advisories daily and have an emergency patching window ready.
- Decommission legacy Ubuntu: Ubuntu 16.04, 18.04, and 20.04 are listed as vulnerable with no fix planned. If any of these run MPTCP in production, migrate or isolate immediately.
- Detection guidance: Monitor kernel logs for panic messages related to TCP socket handling. Configure SIEM alerts for unexpected kernel oops or memory corruption indicators on exposed Linux hosts.
Broader Implications
This vulnerability underscores two trends worth tracking. First, AI-assisted exploit development is demonstrably accelerating the gap between patch availability and weaponization. Anthropic's AI produced a working exploit for a kernel race condition—a class historically difficult to exploit reliably. Second, the MPTCP dependency means organizations that don't even know they have the feature enabled may be exposed. Default configurations matter, and this flaw is a reminder that obscure kernel features can become critical attack surfaces overnight.
Shield53 assesses this vulnerability as high-priority for any organization running exposed Linux infrastructure. The combination of remote unauthenticated access, root-level impact, and AI-proven exploitability elevates this above typical kernel CVEs. Treat the vendor scoring gap with caution—defend against the worst case, not the optimistic one.
