The Model Didn't Find the Bug — It Wrote the Exploit
CVE-2026-98432 is a 9.8 Linux kernel TCP race. Claude built a working remote-root chain for it in under an hour of wall-clock time.
CVE-2026-98432 is a race condition in the Linux kernel's IPv4 TCP connection setup that produces a use-after-free. The kernel CVE team scored it 9.8. The upstream fix landed on October 5 in kernel 6.12.113 and 6.18.56; The Hacker News covered the public disclosure on October 29, ten days after the CVE was published and after Anthropic's companion blog post went up on October 27.
The bug itself is important. The finder story is a little different.
What the fix credit actually says
The kernel fix credits Xinyang Ge of Anthropic with reporting the bug. Anthropic says its Claude model wrote an exploit after "initial prompting." Anthropic doesn't claim Claude discovered the vulnerability — the finding, per the credit line, is Ge's. What's new is the step after the finding.
Per Anthropic's writeup, as relayed by The Hacker News:
- Stage 1 — info leak. The race causes the kernel to send extra bytes back to the attacker over the connection setup. Once a reply contains a kernel address, the attacker has defeated KASLR. Anthropic reports 48 minutes of wall-clock time to leak an address.
- Stage 2 — control. The same bug overflows a buffer and overwrites a callback pointer. That pointer is redirected to a payload already placed in kernel memory, which runs an attacker-chosen command as root. Anthropic reports 5 more minutes from leak to shell.
- Retry economics. A lost race fails only that attempt without crashing the system, so the attacker just tries again. This is the exploit designer's dream property — unlimited free tries on a 9.8 bug.
Anthropic's chosen target was servers with a multipath TCP listener. It did not publish the exploit code, the specific kernel version, or the configuration.
Why the finder's cost matters more than the finder's title
The usual argument against taking frontier-lab exploit demos seriously goes like this: the labs have good models, the labs have good vulnerability researchers, the demos are built on both, so the demos tell you about the labs and not about the broader threat. Fair point when it was true.
What CVE-2026-98432 changes is the step size. Writing an exploit for a Linux kernel TCP use-after-free that includes a KASLR break and a payload-redirect-via-callback-pointer is the sort of task that used to be a named human researcher's months-long project. Not a weekend. Not an afternoon. Anthropic's own numerators — 48 minutes to kernel address, 5 minutes to shell — describe a specific wall-clock-time envelope that any operator with comparable tool access can now reproduce in principle. The report has to be read that way, not as a vendor flex.
Pair this with the October 29 Cipher/EPICS story — Causal Security's commercial agent writing four CVEs (one at 9.8) into industrial control software in under four hours with no human intervention on the initial pass — and the picture is: the frontier-lab unreleased model and the commercial off-the-shelf agent are now both anchors on the same timeline. The step from private-lab demo to in-market tool is getting shorter.
What's still patched and what still isn't
As of the October 29 writeup, the roll looked like this per the distribution advisories cited in the Hacker News article:
- Patched: Upstream (6.12.113, 6.18.56); Oracle Linux and Debian trixie (Oct 16); Debian 12 (Oct 17 / 22); Amazon Linux 2023 kernel6.18 (Oct 24) and kernel6.12 (Oct 29); SUSE Linux Micro 6.2 and openSUSE Leap 16.0.
- In progress: SLES 15 SP7 fix in flight; some Amazon Linux 2023 kernel packages pending; Ubuntu 22.04 / 24.04 / 26.04 / 26.10 unreleased.
- Listed vulnerable, no release: Ubuntu 16.04 / 18.04 / 20.04; RHEL 8 / 9 / 10 and OpenShift 4. RHEL 7 is not affected.
The exposed surface is specifically Linux servers running a multipath TCP listener. That is not every Linux server, but it is far more than zero — multipath TCP is routinely enabled on cellular middleboxes, specific CDN edge configurations, and anywhere path-redundant transport is turned on.
What this means for the enterprise operator
Three things, in order of urgency:
-
Roll the kernel on anything that terminates MPTCP from the public internet. This is the normal patch-rollout response and the only full fix. If your distribution isn't out yet, inventory the MPTCP-enabled hosts and plan the window.
-
Treat "fast kernel exploit from a vendor's AI assistant" as a baseline cost assumption going forward. The industry argument used to be that practical exploitation of a 9.8 kernel race takes weeks of specialist work and therefore buys you weeks of grace between CVE publication and real-world exploitation. That assumption is now visibly weaker. The gap between CVE assignment and working exploit shrinks when the exploit author is a model with no sleep schedule.
-
Watch the behavior at the TCP service boundary. The exploit requires an attacker on the wire to the vulnerable listener, repeatedly, with retries. The individual packet is not malware; the sequence is. A kernel-level listener that starts leaking memory contents, or that spawns a userland process it never did before, is a defender-visible chain regardless of which CVE drove it.
The takeaway
An Anthropic researcher found a 9.8 use-after-free in the Linux kernel's TCP setup. An Anthropic model wrote the exploit. The vendor's own numerators are 48 minutes to kernel address, 5 more to a shell. The patch is out. The patch rollout is not.
The attacker's workbench has a new tool in it. The defender's workbench needs the same step-change — not more pattern files for yesterday's exploits, but the surface that watches what the service process actually does when the race is being driven.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.