CVE-2026-98432: The Linux Kernel Patch Shipped. The Bug Did Not Go Away.
An independent reproduction study of the critical IPv4/TCP use-after-free that Anthropic's Frontier Red Team found and exploited — and how to tell whether your kernel is really patched.
A crafted SYN can race a child-socket teardown so a kernel IPv4-options object is freed while another CPU is still reading it, leaking memory to a remote attacker and, chained, enabling code execution. Upstream fixed it by deferring the free with RCU. But an independent check of the shipped source shows the vulnerable free is still in place at both sites in a stock distribution kernel.
Why it matters
This is a remote, unauthenticated, root-level code execution in the Linux TCP stack — CVSS 9.8. If an attacker can reach port 80/443, the same IPv4 path that serves the web is the one that races. The patch moved the object free to an RCU callback; if your kernel still frees it directly, you are running the vulnerable code.
Key findings
- The fix is not in the stock source The vulnerable free is still present at both free sites in the source tree of a stock distribution kernel — the "shipped patch" defers the free, but the free site in the released source was not moved to the RCU callback.
- Minimal kernels never reproduce A minimal, hand-built kernel reproduces zero times across the full run; a stock distribution kernel reproduces repeatedly — because the race window depends on scheduler and memory-reclaim behaviour, not just the code.
- 71.5 million attempts Across a stock distribution kernel, the race is triggerable over 71.5 million connection attempts — a realistic remote attack surface on any public IPv4 host.
- How to check your kernel The decisive gate is in the disassembly: a vulnerable build calls
kfreeon the IPv4-options object; a fixed build callskvfree_call_rcu. If your build still shows a plainkfree, you are not patched.
Mitigation is a firewall, not a kernel upgrade
- No fixed kernel was shipped at the time of disclosure — an upgrade alone does not close this.
- The base mitigation is a Phase-0 drop on the TCP segment-option bits that trigger the race (nftables/iptables), applied to the inbound 80/443 path.
- Narrower hardening — disabling MPTCP, setting
tcp_migrate_req=0— reduces but does not remove the exposure; they are not the base fix. - Verify with the
kfreevskvfree_call_rcudisassembly gate, not by kernel version number.
What it means for practitioners
Do not assume that a distro "kernel update" means CVE-2026-98432 is fixed. Confirm at the binary level: if your kernel still calls kfree on the IPv4-options object, apply the Phase-0 inbound drop on the trigger bits and keep it until you have verified a build that calls kvfree_call_rcu.
Read the full finding
This page is an ASI Lab news summary of the finding at /CVE-2026-98432/. See that page for the full technical analysis, the disassembly gate, and the verification steps.