CVE-2026-98432: How the Linux Kernel TCP Flaw That Claude Wrote a Root Exploit For Works, and Why Every Distribution Scores It Differently

By Carlos Montiel | Enterprise AI Specialist
Leer en español →
Published: 2026-10-30 | By: Carlos Montiel | Reading time: ~5 minutes
Claude wrote a root exploit for a TCP flaw in the Linux kernel, rated CVSS 9.8 by the kernel team

On October 27, Anthropic reported that an internal research model, after an initial prompt, wrote a Linux kernel exploit that gives a remote, unauthenticated attacker a root shell. The flaw is CVE-2026-98432, and what makes it a good case study is that the same vulnerability gets scores between 5.9 and 9.8 depending on who rates it.

What CVE-2026-98432 Is and How It Works

It's a race condition in TCP connection establishment over IPv4. According to The Hacker News, when the child socket is created, the tcp_v4_syn_recv_sock() function takes the IP options from the pending request socket without copying them. If the child is freed while another CPU is still building a SYNACK for a retransmitted SYN, the kernel reads memory that has already been freed: a use-after-free. IPv6 isn't affected because it does copy the options. The fix, per the description Tenable picks up, uses kfree_rcu() to wait for other CPUs to finish reading before freeing.

The Hacker News states the flaw dates back to Linux 4.4 and that the fix commit was applied on October 5; we couldn't confirm either of those in a second source, so treat them as single-source. Tenable lists the CVE's publication on October 15.

What Anthropic Says (and What It Doesn't)

In its October 27 post, Anthropic's Frontier Red Team writes that, after an initial prompt, Claude produced an exploit that lets a remote attacker send packets to Linux servers with a multipath TCP (MPTCP) listener and run arbitrary commands as root, with no credentials or user interaction. It cites timings: 48 minutes to leak a kernel pointer and another five to reach command execution. It says the patch is upstream and backported to all maintained stable branches.

What isn't stated: Anthropic describes the model only as "an unreleased internal research model"; it gives no name. It also gives no mitigations or CVSS score. The MPTCP vector comes from Anthropic's statements; The Hacker News clarifies that it doesn't appear in the CVE record or in the fix. No production attacks were reported in the sources we reviewed.

Why the Scores Differ So Much

Source Score Kernel CVE team (CNA) 9.8 (AV:N/AC:L/PR:N/UI:N) Tenable 9.8 (critical) Red Hat, current page 8.1 (AV:N/AC:H) "preliminary" Red Hat, per THN at publication "Moderate, 5.9" Oracle and SUSE (per THN) Important, 8.1 Amazon (per THN) Important, 7.0

The key difference is attack complexity: the kernel team assumes AC:L (low) and Red Hat AC:H (high), because the timing race isn't won every time. Red Hat calls its figures preliminary, and its page has changed since THN cited it. The operational moral: a score is an opinion about how hard exploitation is, not about whether an exploit exists; and here Anthropic says a working one does.

What to Do

What It Means for a Company Using AI

Takeaway for your company: the "patch is published, someone turns it into an exploit" cycle can now take less than an hour of compute with a capable model. For anyone running AI infrastructure on Linux, the time between a patch existing and you deploying it is the real risk window. Keep an inventory of which kernel each server runs, automate reboots for kernel patches, and don't let one vendor's low score decide priority on its own.
Carlos Montiel
Enterprise AI Solutions Architect
Specialist in LLMs, Agents, and Orchestration
guatemalia.com/#contacto · info@guatemalia.com

Need to implement AI in your company?

Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.

Contact Carlos Montiel

info@guatemalia.com