Public exploits land for four Linux kernel root flaws

Working exploit code for four Linux kernel flaws is now public, and each one hands an unprivileged user full root on a machine that has not been patched. Security researcher Asim Viladi Oglu Manizada published the exploits and a technical write-up on 18 September, roughly two months after reporting the bugs to the kernel team and after the fixes had already made it into the stable releases. There are no reports of any of the four being used against anyone yet, which is the good news. The less good news is that the underlying mistakes are between 10 and 21 years old, which means a large number of long-lived servers are sitting on code that was written before any of this was understood.

What the four flaws are

The quartet is DirtyAH6 (CVE-2026-80844) in the IPsec code that handles the IPv6 Authentication Header, TUNderflow (CVE-2026-81000) in the TUN and TAP virtual network devices, PPPoEject (CVE-2026-68121) in PPP over Ethernet, and DiagSpill (CVE-2026-74469) in the reporting code for SCTP. All four are memory-safety bugs in kernel networking code that let an attacker corrupt kernel memory and turn that corruption into a root shell.

Three of them, DirtyAH6, TUNderflow and PPPoEject, are only reachable by an ordinary user when unprivileged user namespaces are switched on, a feature most mainstream distributions enable by default because containers and sandboxes lean on it. DiagSpill is the odd one out: it needs no special privileges and no user namespaces at all, as long as the SCTP module is available. In both cases the attacker already needs a foothold, which is what makes this a privilege escalation problem rather than a break-in, and the stakes are highest on shared and multi-tenant systems where a low-privileged account is easy to come by.

The fixes are already out

Manizada reported the bugs to security@kernel.org in mid-July, and the fixes landed in the stable trees over the following weeks. The first stable releases carrying the complete set are 5.10.270, 5.15.221, 6.1.188, 6.6.157, 6.12.109, 6.18.50 and 7.2.4. Those are the kernel project's own version numbers, and almost nobody runs them: Debian, Ubuntu, Red Hat and SUSE each ship their own numbering and back-port fixes on their own schedule, so the useful check is your distribution's security advisory rather than matching a number from the list above.

What to do if you cannot patch today

Turning off unprivileged user namespaces removes the everyday path to three of the four flaws, though it does not stop DiagSpill and will break some container setups. Turning off AH6, TUN and TAP, PPPoE and SCTP closes the affected code entirely, at the cost of whatever depended on it. Manizada's own advice is to patch rather than to switch features off, on the reasonable grounds that other routes into the same bugs may exist and that a disabled feature is easy to forget about. The exploits themselves are tuned to specific kernel builds and can crash the machine, so they belong on isolated test systems.

AI is doing the finding now

The four flaws were found with an AI-assisted harness that builds a map of how the kernel handles memory and reasons about its layout, and the fix commit for DirtyAH6 carries an "Assisted-by" line crediting that tooling. It is one entry in a busy 2026 for kernel privilege escalation, several of them found with large language models, and Manizada said this batch likely closes the public phase of his AI-assisted bug hunt. One exploit also reuses a technique from Dirty Frag, a kernel root flaw disclosed in May by another researcher.

Our opinion

The interesting number here is not four, it is 21. A memory-safety bug that has survived two decades in the same networking subsystem is not a sign of careless maintainers so much as proof of how much kernel code was written before anyone was paid to look at it this closely, and the standard remedy, a two-day embargo and a stable-tree back-port, only reaches people who take updates. What has genuinely changed is the economics of finding them. A harness that reasons geometrically about memory layout and a model that reads the result is how one researcher produced four root exploits and a 2,000-word write-up on his own, and the same leverage is available to anyone who wants it. The sobering corollary is that the widely repeated hardening advice, disable unprivileged user namespaces, is also the advice that breaks the container platforms most European infrastructure has just finished migrating onto. Patching is the dull answer and, for once, plainly the right one.