Exposure audits, hunt queries and a validated interim mitigation for CVE-2026-98432, the use-after-free in the Linux IPv4 TCP handshake. Read-only scripts with a self-test. From GarphenGate Security Research.
Exposure checks, hunt queries and the interim firewall mitigation for CVE-2026-98432, the use-after-free in the Linux IPv4 TCP handshake, from GarphenGate Security Research.
The full write-up, with our lab evidence and the reasoning behind every check here, is on our blog: CVE-2026-98432: A Defender's Field Guide to the Linux TCP Use-After-Free.
Patching and rebooting is the only complete fix. Everything in this repository helps you find exposed hosts, work out which ones to do first, and see attempts while you wait. None of it replaces the kernel update.
The bug is in code that runs when a Linux server accepts an IPv4 TCP connection, before any application sees the connection. There is no vulnerable package to remove, no application setting to change, and no WAF in front of it unless you put one there deliberately. Your web server, your SSH daemon, your database and your Kubernetes nodes are all the same host as far as this bug is concerned.
It also reaches back to Linux 4.4, so "we run something recent" tells you nothing.
Upstream stable releases carrying Eric Dumazet's fix (kfree → kfree_rcu, commit e3f33b0):
| Branch | Fixed in | Released | Status |
|---|---|---|---|
| 7.3 | 7.3 | Oct 18, 2026 | stable |
| 7.2 | 7.2.10 | Oct 11, 2026 | branch EOL (last release 7.2.15, Nov 1) |
| 6.18 | 6.18.56 | Oct 11, 2026 | longterm |
| 6.12 | 6.12.113 | Oct 15, 2026 | longterm |
| 6.6 | 6.6.159 | Oct 15, 2026 | longterm |
| 6.1 | 6.1.190 | Oct 15, 2026 | longterm |
| 5.15 | 5.15.223 | Oct 15, 2026 | longterm |
| 5.10 | 5.10.272 | Oct 15, 2026 | longterm |
| 4.4 – 5.4 | n/a | n/a | upstream EOL; distro fixes only |
Compare against your distribution's tracker, not this table. Distro kernels carry the fix backported under their own version numbers, so 6.8.0-149.149 can be fixed while looking older than 6.12. The audit script here tells you which tracker to read; it cannot tell you what the backport did.
Reboot after installing. A fixed kernel on disk does nothing until the machine boots into it.
| Path | What it does |
|---|---|
audit/kernel-exposure.sh | Running vs installed kernel, upstream floor comparison, pending-reboot detection, and which tracker applies to you |
audit/mptcp-exposure.sh | Finds MPTCP listeners (the precondition Anthropic's working exploit needs) and the Go 1.24+ processes likely to own them |
hunt/ip-options-traffic.sh | Hunts captures and interface counters for IPv4 packets with options, the attacker's only input |
hunt/uaf-signatures.sh | Scans dmesg/journal for use-after-free reports naming the functions in this bug |
hunt/handshake-anomalies.sh | SYN-queue and handshake counter anomalies that accompany sustained racing |
mitigate/nftables-drop-ipopts.sh | Applies, verifies and removes the nftables rule; refuses to touch a host using CIPSO |
mitigate/iptables-u32-drop-ipopts.sh | Same rule for iptables shops, with an xt_u32 availability check |
docs/hardening-checklist.md | Interim hardening while patching waits, and what each measure does and does not buy |
Every script is read-only except the two in mitigate/, which say exactly what they will change and ask first. They need only bash, grep, awk and sed. Exit codes: 0 clean, 1 findings, 2 usage error or cannot determine.
git clone https://github.com/GarphenGate/cve-2026-98432-defender-toolkit.git
cd cve-2026-98432-defender-toolkit
bash tests/run-tests.sh # self-test on synthetic data
# 1. Where does this host sit?
bash audit/kernel-exposure.sh
# 2. Is it one of the hosts an exploit actually needs?
bash audit/mptcp-exposure.sh
# 3. Buy time if you cannot patch today
bash mitigate/nftables-drop-ipopts.sh --show # preview, changes nothing
bash mitigate/nftables-drop-ipopts.sh --apply
# 4. Has anything already tried?
bash hunt/ip-options-traffic.sh
bash hunt/uaf-signatures.sh
Run these on container hosts and Kubernetes nodes, not inside containers. A container reports the host's kernel and has no visibility into it.
If the host accepts IPv4 TCP connections from anywhere you do not fully control, and its kernel predates the floors above, yes. Everything else in this repository is about ordering the work and seeing the attempts.
We reproduced the bug in an isolated lab and confirmed the use-after-free, the exposed code path, and the mitigation. We did not build a working remote root exploit, and nothing here demonstrates one. Anthropic reported that its Claude model built such an exploit against servers with an MPTCP listener; it has not published it. What that means for your patch ordering is in the article.
MIT licensed. See LICENSE.