News › security
By The Linux Camp News, October 31, 2026
Anthropic's October 27 report describes a remote exploit for CVE-2026-98432. The fix delays freeing IPv4 options; operators need to verify distribution backports and the running kernel, not just its base version.
Anthropic publicly reported a Linux networking exploit on October 27, 2026, demonstrating unauthenticated remote root command execution against servers with multipath TCP listeners, according to LWN and Anthropic. Anthropic identifies the flaw as CVE-2026-98432 and, in its report retrieved October 31, says fixes reached upstream and all maintained stable kernel branches. October 27 is the research-report date, not the discovery date: the official CVE record was already published October 15. 1 3 4 5
The defect is an ownership-and-lifetime race involving IPv4 options. During connection establishment, tcp_v4_syn_recv_sock() transfers the options pointer from the pending request to the new child socket without copying the options. Meanwhile, another CPU can answer a retransmitted SYN using that same memory. If the child socket is destroyed before the reply is built, its destructor immediately frees the options. The reply-building code can then read freed memory in ip_options_build(): a use-after-free, rather than simply two connections being established simultaneously. 2 4
The fix changes memory reclamation, not the handshake protocol. Readers already use RCU, the kernel mechanism that lets existing readers finish before shared memory is reclaimed. Replacing immediate freeing with kfree_rcu() in both inet_sock_destruct() and tcp_v4_reqsk_destructor() makes reclamation wait for an RCU grace period. That closes the lifetime gap without requiring the readers to stop using the options early. 2 1
Anthropic describes two exploitation stages: leaking a kernel pointer to defeat kernel address-space randomization, then corrupting a callback pointer to redirect execution. Its proof of concept reportedly needed 48 minutes for the leak and another five minutes for shell-command execution. Anthropic says failed races only failed the handshake in this demonstration, permitting retries. Those timings and behavior describe its experiment, not a universal exploit duration or assurance that other attempts cannot crash a host. 3 1 5
Keep the demonstrated target separate from the underlying defect. Anthropic demonstrates exploitation through multipath TCP listeners; the maintainer describes a fault in shared IPv4 TCP code. The packet does not establish equivalent command execution against ordinary TCP listeners. IPv6 avoids this particular child-socket race because it copies the options, but the maintainer explicitly identifies a separate IPv6 request-migration issue involving CALIPSO options. That is not a blanket IPv6 safety verdict. 3 2
For triage, prioritize checking reachable multipath TCP listeners and ask the distribution vendor which kernel package carries this fix. The reported remote, pre-authentication impact makes those listeners an immediate review target. Treat package installation and activation as separate checks: verify the running kernel against the vendor's patched build, not merely the presence of an installed package. Because fixes are backported, an older base-version number alone is not a patch verdict. The supplied sources do not establish distribution package availability. 3 1 5
The official CVE record retrieved October 31, revised October 30, supplies upstream release boundaries missing from the research narrative. Examples of fixed stable releases are 6.1.190, 6.6.159 and 6.12.113, with subsequent releases on those respective branches marked unaffected. These are upstream reference points, not a complete fixed-version list or substitutes for distribution backport confirmation. 5
As retrieved October 31, the official CVE record assigns CVSS 3.1 severity of 9.8, Critical. That score is distinct from Anthropic's research demonstration and is not evidence of observed attacks. The supplied CISA lookup, retrieved October 31 from the October 30 catalog snapshot, does not list this CVE. Absence from KEV does not establish absence of exploitation; the packet provides no confirmed in-the-wild attack report. 5 6