Captain’s log · Stardate 26310.6 · Engineering
remote root and record domains
An AI just found a remote root exploid; how I deal with the new threat of AI. People are buying record numbers of domains for AI projects. Here is what it takes to keep them alive
On October 27, Anthropic reported that one of its unreleased research models wrote a working remote root exploit for a bug in the Linux kernel (CVE-2026-98432).
The bug is a race condition in how Linux sets up IPv4 TCP connections when a server is listening for Multipath TCP (MPTCP). According to Anthropic's write-up, the exploit took 53 minutes to run (48 minutes to leak a kernel pointer, and 5 more to get a root command). That 53 minutes is how long the attack took against the test server, not how long the AI took to write the code.
When a story like this comes out, there is an obvious irony for me. The company whose AI just built a root exploit for Linux is the exact same company whose AI I use every day to help run my Linux servers.
I have a local server inside our motorhome network where an instance of Claude Code (we call him Maquinista) has SSH key access to our server fleet. Instead of logging into a dozen boxes by hand to see if we were exposed, I asked Maquinista to check all of our AlmaLinux servers for me.
He checked the kernel settings and active listeners across the fleet right away. Every single server had MPTCP turned off, because that is the default on AlmaLinux and we never had a reason to turn it on. With no MPTCP listener running, the remote exploit has nothing to talk to.
Still, a kernel bug rated up to 9.8 is not something you leave unpatched just because a feature is off by default. On October 30, the fixed kernels were still in the AlmaLinux testing repository, and a fix published, a package shipped, and a server rebooted are three different days.
So I asked Maquinista to set up a watcher (kernel-cve-watch) to check the production repositories, install the patched kernels as soon as they arrived, and email me so I could handle the reboots myself.
This afternoon at 12:38 PM, I got the email I was waiting for:
[kernel-cve-watch] CVE-2026-98432: 5 installed, 0 need a look, 7 host(s) left
Fixed kernel installs (NOT rebooted, every box still runs the old kernel until its reboot):
OK dq05: RESULT=installed new=6.12.0-211.72.1.el10_2.x86_64 default=6.12.0-211.72.1.el10_2.x86_64 running=6.12.0-211.69.1.el10_2.x86_64
OK mudanca: RESULT=installed new=5.14.0-687.64.1.el9_8.x86_64 default=5.14.0-687.64.1.el9_8.x86_64 running=5.14.0-687.61.1.el9_8.x86_64
OK ns1: RESULT=installed new=5.14.0-687.64.1.el9_8.x86_64 default=5.14.0-687.64.1.el9_8.x86_64 running=5.14.0-687.61.1.el9_8.x86_64
OK panel: RESULT=installed new=5.14.0-687.64.1.el9_8.x86_64 default=5.14.0-687.64.1.el9_8.x86_64 running=5.14.0-687.61.1.el9_8.x86_64
OK directadmin: RESULT=installed new=5.14.0-687.64.1.el9_8.x86_64 default=5.14.0-687.64.1.el9_8.x86_64 running=5.14.0-687.61.1.el9_8.x86_64
The first five AlmaLinux 9 and 10 machines (dq05, mudanca, ns1, panel, and directadmin) already have the new kernel installed and set as default, and the watcher is waiting on the remaining seven hosts (plus our failsafe box, which manages its own kernel). Notice the first line of that email: Maquinista installs the kernel, checks that the DKMS modules built cleanly, and sets the default boot entry, but he does not reboot the servers. I still make that call myself.
Le and I just celebrated 30 years of RogerLe.com on October 31. After three decades of managing servers, two lessons from this week stand out:
- Boring defaults save you. Keeping AlmaLinux as close to default as possible meant MPTCP was off everywhere before we even looked.
- Use the tool that caused the problem to help fix it. If an AI model in a lab can write a kernel exploit, you shouldn't be checking twelve servers by hand. Let your own AI check the settings, watch the repos, and stage the updates so all you have to do is pick when to reboot.
People are buying record numbers of domains for AI projects.
On October 27, OVHcloud reported that its daily domain sales hit record volumes, and gave credit to "vibe coding" and AI site builders. Verisign said something similar the week before, pointing to AI as one of the reasons behind 13.1 million new .com and .net registrations between July and September.
There is an odd detail in OVHcloud's report, though. Even with record daily domain sales, their web hosting revenue grew only 3.2% over the past year (to 199.7 million euros).
Those figures come from the companies' own explanations rather than an independent study, so take the exact cause with a grain of salt. Still, as someone who just celebrated 30 years of running RogerLe.com on October 31, my guess at why domain sales are jumping while hosting barely moves is pretty simple. Buying a domain takes two minutes and ten dollars. Turning that domain into a live website that stays online takes real infrastructure work.
I know the feeling of registering new domains for AI projects, because I bought three of them myself recently.
In our case, we do not call it vibe coding, we call it built by the crew. Different members of my AI crew built three separate sites for things I personally care about:
morais.news, my personal news site built to track the exact topics and news I want to read every day (built by Claude).exoplanetledger.org, a catalog of exoplanet data pulled from the NASA Exoplanet Archive (built by Gemini 4).athleticstats.org, a site tracking the top 50 athletes in the world for each Olympic event, plus the top 500 performers overall (built by GLM 5.3).
Having the crew write the PHP, JavaScript, and database queries is the fast part. If I stopped there, I would just be another person with three domain names and some code sitting in a folder.
To make morais.news, exoplanetledger.org, and athleticstats.org real websites that anyone can visit, the old hosting rules from 1996 still apply:
- DNS and SSL. Somebody has to point the domain records properly, issue the certificates, and make sure they renew on time.
- The web server stack. The code needs a properly configured
nginxand PHP-FPM setup, with rules that block bad requests before they touch the application. - Database and cron jobs. Data from places like the NASA Exoplanet Archive or news feeds has to be fetched, cleaned, and indexed without filling up the disk or hanging the server.
- Backups and security. As we saw with the PHP and Linux kernel bugs over the last two weeks, an unmonitored server gets probed within days.
My guess is that a lot of those 13.1 million new domains were bought on a Saturday afternoon while someone was excited about what an AI wrote on their laptop, and abandoned on Sunday when they hit DNS, email deliverability, and server setup.
Building with an AI crew is the most fun I have had in programming in years. Just remember that after the AI finishes writing the code, somebody still has to run the server.