AnalysisNov 10, 20266 min read6 sources

Anthropic's Linux remote-root claim holds up; CVE-2026-98432 is a core TCP bug, not just an MPTCP one

Anthropic's post backs the zero-click remote root on MPTCP listeners in about 53 minutes. Red Hat now rates the IPv4 TCP use-after-free Critical 9.8, and Go tells users to disable its default-on MPTCP listeners. A patched kernel plus a reboot is the only real fix.

In one minute

Anthropic's post backs the zero-click remote root on MPTCP listeners in about 53 minutes. Red Hat now rates the IPv4 TCP use-after-free Critical 9.8, and Go tells users to disable its default-on MPTCP listeners.

On October 27, 2026, Anthropic's Frontier Red Team published "Discovering vulnerabilities with Claude." One section describes zero-click remote code execution in Linux that an unreleased internal research model wrote "after initial prompting." The bug is CVE-2026-98432, a race-condition use-after-free in the kernel's IPv4 TCP connection setup. The upstream fix is "ipv4: free inet_opt and ireq_opt after an RCU grace period." The demonstrated exploit targets servers with a Multipath TCP (MPTCP) listener.

This matters to builders for two reasons. First, Go 1.24 and later open MPTCP listeners by default on Linux, so ordinary Go services can expose exactly the attack surface the exploit used. Second, Red Hat first scored the bug as a lower-severity, possible denial-of-service issue, then reissued it as Critical. At the end of October, fixed kernels were still missing for RHEL and Ubuntu.

What we checked

  • Does Anthropic's full public post substantiate the Linux CVE, zero-click exploit conditions, and claimed time to remote root access? partly confirmed. Yes on the CVE number, the zero-click unauthenticated conditions (MPTCP listener) and the timing (48 minutes to leak a kernel pointer, about 5 more to root commands). But an unreleased internal model wrote the exploit after initial prompting, the post gives no kernel versions or explanation of MPTCP's role, and the exploit is unpublished and has not been independently reproduced. [1]
  • (Added) Is the vulnerability MPTCP-specific, so that disabling MPTCP is enough? refuted. Go, AlmaLinux and CIQ (as reported by BleepingComputer) all say the bug is in IPv4 TCP connection setup and not MPTCP-only. Disabling MPTCP and filtering IPv4 options are partial mitigations; a patched kernel plus a reboot is the fix. [5][2][3]
  • (Added) Did Red Hat actually raise its rating to Critical 9.8? confirmed. Red Hat's CVE page shows Critical with CVSS 9.8, and RHSB-2026-016 is dated October 30. BleepingComputer reports the earlier 5.9 denial-of-service framing. [6][4][3]

What we found

  • Anthropic's post backs the core claim. It names CVE-2026-98432, calls it a race-induced use-after-free in the networking stack, and says the proof of concept leaked a kernel pointer (defeating KASLR) in 48 minutes, then reached root command execution about five minutes later. That needs no credentials or user interaction against a server with an MPTCP listener. The bug was patched upstream before the post went out. The post does not name affected kernel versions, does not explain why the MPTCP listener matters, and does not release the exploit. [1]
  • AlmaLinux pins down the trigger: the bug needs an incoming TCP connection request carrying IPv4 options. The upstream fix is commit e3f33b0a1d89. MPTCP is off by default on AlmaLinux. Its suggested stopgaps, an nftables drop rule and disabling MPTCP, are explicitly not fixes. [2]
  • As of October 30, BleepingComputer reported no confirmed exploitation in the wild. Ubuntu 22.04 through 26.10 were still unpatched, while Debian 12 and 13 had fixed kernels. Red Hat's bulletin named no fixed RHEL kernel yet. [3][4]

What it means

Treat this as a patch-now kernel bug for any internet-facing Linux host, MPTCP or not. Check the running kernel against your distro's fixed build, and reboot. Installing a package does not change the running kernel. Until fixed kernels land (RHEL and Ubuntu were still pending on October 30), deploy the nftables rule that drops inbound IPv4 TCP packets with IP options (ip hdrlength > 5). Test it first, because it can break rare legitimate traffic. On hosts that don't need MPTCP, also set net.mptcp.enabled=0. For Go services built with 1.24 or later, add godebug multipathtcp=0 to go.mod or set GODEBUG at runtime, and confirm with go version -m. Libraries don't inherit the main module's setting.

The bigger signal for AI builders: a frontier lab's internal model turned a kernel race into a reliable remote root in under an hour of runtime. A downstream vendor first scored the same bug as a moderate DoS and changed its rating only after the AI-built exploit was disclosed. Expect "complex race, probably just DoS" ratings to get less trust, and severity scores to change faster. Budget for faster kernel patching, and watch default-on features like Go's MPTCP listeners that quietly widen your attack surface.

Competing readings

  • A genuine, widely affecting kernel RCE: the bug sits in core IPv4 TCP code, Anthropic's PoC proves remote root is practical, and vendors (Red Hat, Go) are right to treat it as urgent. MPTCP is only how the demonstrated exploit gets in (likely, about 70%). What would tell it apart: Red Hat moved from a DoS framing to Critical and says only kernel updates fully fix it [6][4][3]; Anthropic describes zero-click root in about 53 minutes on a patched-upstream bug [1]; Go and AlmaLinux both say the bug is not MPTCP-only and require a kernel patch [5][2].
  • Real bug, but practical risk is narrower than 9.8 suggests: exploitation as shown needs an MPTCP listener, IPv4 options and winning a race. The urgency comes mainly from the AI-exploit framing, and the exploit is unpublished and not independently reproduced (unlikely, about 27%). What would tell it apart: Trigger requires IPv4 options; MPTCP is off by default on EL distros; there is no confirmed in-the-wild exploitation [2][3]; Red Hat originally treated it as a potential DoS [3].
  • Anthropic's claim is overstated or unreproducible in practice, and the severity upgrade is reputational rather than technical (very unlikely, about 5%). What would tell it apart: Red Hat's own bulletin still cites Anthropic's post, and Red Hat now rates the bug Critical [6].

Base rate

The New Runtime archive holds only 4 earlier research stories about companies, models or projects from the last 12 months that are at least 3 months old: too few (minimum 5) for a base rate. The archive has only 4 comparable research stories (3 matches), which is below the minimum, so the base rate tells us nothing here. This analysis rests on the primary sources instead.

What would change this view

  • Independent researchers reproduce remote code execution without MPTCP, or a public exploit appears (signal): Risk is broader than the MPTCP-listener scenario, and the nftables and MPTCP mitigations are much weaker than hoped.
  • Red Hat or the kernel CNA lowers the score again after analysis, or several vendors fail to reproduce the RCE (signal): The practical exploit is harder than Anthropic's PoC implies.
  • Go accepts issue #82717 and ships MPTCP off by default (deprecation signal): The Go team now sees default-on MPTCP listeners as a lasting attack surface, not a one-off.

Still open

  • Why does Anthropic's exploit depend on an MPTCP listener when the bug sits in IPv4 TCP code, and can it be adapted to plain TCP listeners?
  • Which exact kernel versions are affected, and which distro builds are fixed? Anthropic gives neither, and vendor status was changing at the end of October.
  • Has any independent party reproduced the remote code execution? Reports that CloudLinux could not reproduce it surfaced only in search results and were not verified here.
  • Will Go change its default MPTCP listener behavior (issue #82717)?

Evidence / sources

  1. 1anthropic.com/research/discovering-vulnerabilities-with-claude
  2. 2almalinux.org/blog/2026-10-30-cve-2026-98432
  3. 3bleepingcomputer.com/news/security/critical-linux-kernel-flaw-allows-remote-root-attacks-on-mptcp-servers
  4. 4access.redhat.com/security/vulnerabilities/RHSB-2026-016
  5. 5groups.google.com/g/golang-announce/c/PCoQcdxoScw/m/aD-MgXEvAAAJ
  6. 6access.redhat.com/security/cve/cve-2026-98432
All Analysis

Next reads

Read next

Open the graph
  1. 01topicAnthropicExplore the anthropic topic hub.
  2. 02archiveAnalysisOpen the latest authorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract