19 SEP 2026 — Working exploits for four Linux kernel flaws went public at 06:00 UTC on 18 September, each one turning an ordinary local account into root.
The bugs they attack are between 10 and 21 years old. They were found this year by a researcher using AI tooling, and one of the kernel's own fixes credits that tooling in its commit trailer.
What the four flaws are
All four are memory-safety faults in the kernel's networking code, and all four were reported to the kernel security team in mid-July. Asim Viladi Oglu Manizada published the write-up and exploit code after a coordinated hold with Linux distributions so fixes could ship first.
DirtyAH6, CVE-2026-80844, is in the IPsec authentication header path for IPv6. PPPoEject, CVE-2026-68121, is in PPPoE. TUNderflow, CVE-2026-81000, is in the TUN and TAP virtual network devices. DiagSpill, CVE-2026-74469, is in SCTP diagnostics.
DirtyAH6 shows the shape of the class. A routing-header field called segments_left is trusted to be no larger than the number of addresses the header actually describes. That assumption holds for ordinary traffic and fails for raw IPv6 packets where the sender supplies the header. A header describing one address can declare 255 segments, sending an address pointer 4,064 bytes backwards into a memory move that was never meant to reach there.
The one that needs no namespace
Three of the four require unprivileged user namespaces, a feature many hardened builds disable precisely because it widens what a local account can reach.
DiagSpill does not. It needs only that SCTP is available, with no special privileges at all. DiagSpill is therefore the one to check first on systems where user namespaces have been disabled.
Fixes are in stable releases 5.10.270, 5.15.221, 6.1.188, 6.6.157, 6.12.109, 6.18.50 and 7.2.4. Distribution timelines vary, and those are what matter more than the upstream number.
How they were found
Manizada describes an AI-assisted method rather than a fuzzer: graph-based tracking of security-relevant kernel objects and their properties, with agents reasoning about the shape of memory. The same approach produced his earlier CIFSwitch and OVSwrap work.
The distinction matters. Fuzzing finds crashes and leaves a human to work out whether a crash is exploitable. A method that models memory layout aims at the harder question directly, and four working local-root exploits are a strong answer.
The ages of the faults are the uncomfortable part. They sat in widely reviewed networking code for a decade or more, missed by years of scrutiny from people who knew what they were looking for. Something that reads code differently found them in months.
A credit line in the kernel
The fix for DirtyAH6 carries an Assisted-by trailer crediting the researcher's AI tooling, according to reporting on the disclosure. Kernel commit trailers are a formal record of who contributed what, and a tool appearing in that record is a small but real change in how the work is accounted for.
We have not read the commit ourselves — kernel.org refused our automated requests — so we describe the trailer rather than quote it. The precise wording is worth checking before anyone builds an argument on it.
Manizada says this work "likely concludes the AI-assisted vulnerability hunting experiment" in its public form for a while. The method is demonstrated; its author is stepping back from publishing it.
What defenders should check
No attacks have been reported. The exploits are tuned to particular kernel builds and can crash a machine, and their author says to run them only on dedicated virtual machines or throwaway hosts.
Published exploit code shortens the distance between a patch note and a working attack, and local privilege escalation turns a foothold into a compromised host. We covered earlier Linux kernel flaws reaching CISA's exploited catalogue last month; that is the pattern these would follow if anyone picks them up.
The practical question is whether SCTP is loaded on machines that do not need it. DiagSpill needs nothing else. An unused protocol module is the cheapest thing on this list to remove.
View author profile → · Editorial policy
