What happened
CVE-2026-98432 is a use-after-free in the IPv4 TCP handshake of the Linux kernel. It has been there since Linux 4.4 (2016). The Linux kernel CNA published it on 15 October 2026 and scored it CVSS 3.1 9.8 (network, no privileges, no user interaction).
The bug got wide attention on 27 October, when Anthropic's Frontier Red Team described a working exploit: "after initial prompting, Claude wrote an exploit that allows a remote attacker to send packets to Linux servers with a multipath TCP listener to run arbitrary commands as root". The post says the work used an unreleased internal research model. The kernel fix credits Xinyang Ge, an Anthropic researcher, as the reporter. Anthropic has not published the exploit.
No exploitation in the wild has been reported, and the CVE is not in CISA's KEV catalog.
The bug
When a client opens a TCP connection, the kernel keeps a small "request socket" for the pending handshake. If the client's SYN carries IPv4 options (Record Route or Timestamp, for example), those options are stored with that request.
- When the handshake completes,
tcp_v4_syn_recv_sock()hands the options to the new socket without copying them. - Since Linux 4.4, request sockets are handled without the listener lock, so another CPU can still be using those options, for example to build a SYN-ACK for a retransmitted SYN.
- When the socket goes away, the options are freed immediately, while that other CPU may still be reading them.
The attacker needs nothing but network access: one SYN with IPv4 options to any listening IPv4 TCP port. IPv6 is not affected by this path, because tcp_v6_syn_recv_sock() copies the options. A second, similar path exists through request socket migration (net.ipv4.tcp_migrate_req=1 or a BPF migration program), which is off by default.
The fix, written by Eric Dumazet, frees the options only after an RCU grace period, once no CPU can still be reading them (commit "ipv4: free inet_opt and ireq_opt after an RCU grace period").
Anthropic's exploit, and the MPTCP question
Anthropic's proof of concept works in two stages: it first leaks a kernel pointer remotely to defeat KASLR, then overwrites a callback pointer to take control. In Anthropic's words, "it takes 48 minutes to leak a kernel pointer and another five minutes to reach the shell command execution". A lost race only fails one handshake and does not crash the machine, so the attacker can retry until it works.
The published exploit needs a Multipath TCP (MPTCP) listener on the target. That narrows the demonstrated attack, but not the bug:
- the CVE record and the fix do not mention MPTCP;
- the Go Security Team writes that "the underlying bug is not specific to MPTCP";
- Red Hat says MPTCP is not required for the flaw.
MPTCP is off on most Linux services, with one large exception: since Go 1.24, Go programs enable MPTCP by default for listening sockets on Linux. Any Go web server, proxy or agent built with a recent toolchain can therefore expose the exact setup Anthropic exploited. Marcus Hutchins counted about 2.1 million servers reachable over MPTCP on port 443 and, according to Aikido, reported MPTCP probing on that port from 2 November.
Not every vendor agrees on how hard exploitation is. Oracle and SUSE score it 8.1 (high attack complexity), and CloudLinux could not reproduce the result, according to Aikido. Red Hat rates it Critical for RHEL.
Fixed kernels
Upstream stable fixes, released between 11 and 18 October:
| Branch | Fixed in |
|---|---|
| 7.3 | 7.3 |
| 7.2 | 7.2.10 |
| 6.18 | 6.18.56 |
| 6.12 | 6.12.113 |
| 6.6 | 6.6.159 |
| 6.1 | 6.1.190 |
| 5.15 | 5.15.223 |
| 5.10 | 5.10.272 |
Branches 4.4 to 5.4 are end of life and get no fix.
Distribution status on 4 November:
- Debian: fixed (Debian 13 in 6.12.113-1, DSA-6568-1; Debian 12 in 6.1.190-1, DLA-4841-1).
- RHEL: fixed on 1 November for 8.10 (RHSA-2026:93414), 9.8 (RHSA-2026:93412) and 10.2 (RHSA-2026:93413), then for several extended-support streams. OpenShift and some real-time kernels were still pending.
- AlmaLinux and Oracle Linux: fixed kernels released.
- Amazon Linux 2 and 2023: fixed between 24 and 30 October.
- SUSE: fixed for SLES 16.0 and SLES 15 SP4 to SP7 (default kernel); check SUSE's advisories for other flavors.
- Ubuntu: fixes for 22.04, 24.04 and 26.04 were still in testing (
-proposed). 20.04 and older need Ubuntu Pro.
These lists move daily. Check your vendor's advisory for your exact release.
What to do
- Install the patched kernel and reboot. That is the only complete fix. Confirm with
uname -rafter the reboot. - Find MPTCP listeners, the setup the public research targets:
sysctl net.mptcp.enabled sudo ss -Mlnp - Until you can reboot, reduce exposure:
sysctl -w net.mptcp.enabled=0(and a file in/etc/sysctl.d/to keep it), then restart the services, since existing MPTCP sockets stay open. This removes the exploit's prerequisite, not the bug.- For Go services,
GODEBUG=multipathtcp=0in the environment, orgodebug multipathtcp=0ingo.mod. The Go team calls this per-program approach "a fragile mitigation". - Drop IPv4 TCP packets that carry IP options, which legitimate traffic almost never uses:
The rule lives in memory only and is lost on reboot. Do not use it on networks that rely on CIPSO/NetLabel labels.sudo nft add table inet cve98432 sudo nft add chain inet cve98432 in '{ type filter hook prerouting priority -300; }' sudo nft add rule inet cve98432 in meta l4proto tcp ip hdrlength '>' 5 counter drop - Keep
net.ipv4.tcp_migrate_reqat 0.
- Prioritize internet-facing hosts: load balancers, reverse proxies, VPN gateways, Kubernetes nodes and anything running Go services.
- Watch for kernel crashes mentioning
ip_options_build,ip_build_and_send_pktorskb_under_panic. No indicators of compromise have been published.