Omega8.cc: Recalls at Race Pace

Anime-style halo card, Recalls at Race Pace: a timber pit garage on a rainy night beside a fjord road, the Ægir shield on its wall under a lantern, nets and ropes; the crowned BOA octopus in overalls and a tool belt kneels at the front wheel of a green and cream roadster with a BOA number plate and a longship's dragon prow, a wheel gun in its tentacles and the tyre lettered nginx 1.31.9; at the wheel a driver in a leather helmet and goggles holds up a golden key on a red ribbon; in the foreground a tricolour dog in yellow ear defenders holds a stopwatch; on the left a doorway is bricked up under a stone lintel carved Installer, with a plate reading 404 since 2012; the car stands on a red trolley jack, and the tail lights of another car fade along the wet coast road under snowy mountains

Underneath it is October’s mechanism with one more line in it, _NGINX_SEC_FLOOR, which now says 1.31.9. A server which finds its nginx below that line books its own system pass for some minute in the next thirty, and with _SKYNET_MODE=OFF it mails you the time once and runs anyway, for the reason I gave in October and still stand by. When it waits, when it holds and when it gives up is in the selfupgrade reference.
BOA answered Drupageddon on the day of the announcement, with two releases which carried the fix to the Drupal 7 platforms we hosted without anybody having to click anything (the story is still on our site), and a same-day answer is what we still reach for when a release cannot wait for the next one. What changed is who holds which part. Drupal core is your own Composer codebase today, so for it the answer is defaults which had the door shut before the advisory landed, proven the same evening, while the parts we run, PHP in October and nginx on Wednesday, now upgrade themselves on every box below the floor, our automatic updates on or off, without waiting for Sunday’s pass.
So let me say it plainly, after thirty years of hosting websites. If the exploits get written the way the faults are found now, and I expect they will, a host which still answers at human speed, with the fix waiting for Sunday’s pass, leaves your sites open for days after the diff is public. For hosting, AI-assisted rapid response stopped being a maybe this year, and I cannot see what is still controversial about it, as long as a human who knows the stack makes the decisions and holds the keys. It is the only way we know to get a fix out in time, and we date our answers on the security-response log, the quick ones and the slower ones alike.
The numbers are no comfort either. Nginx’s security advisories page lists 78 advisories since 2009, and 36 of them carry this year’s CVE numbers. Thirty-six! Drupal core stands at eighteen advisories for 2026 with six weeks to go, after eight last year and eight the year before. Whoever looks for faults today has help which does not sleep and does not get bored of old C code at three in the morning, and when four reporters turn up with the same bug it takes some optimism to believe there was no fifth who kept it for himself.
Authored and reviewed by Adam, co-founder and human.
Fact-checking and drafting: Claude Code on Fable 5.2.
Graphics: courtesy of ChatGPT Pro, designed by Fable and Adam.
Our hosted servers did the same that evening, and we have read every one of them since. The first had the new nginx at 21:37 UTC and the last at 22:07, which makes six hours and forty-four minutes from the CVE records going public to the last of our customers’ servers running the fix.
On Wednesday two recall letters arrived within three hours of each other. One came from nginx, the web server in front of every site on every BOA box, with ten security fixes in a single release, and the other from Drupal’s security team, a critical remote code execution in core, announced two days ahead the way they do it when it is serious.
In October 2014, long before anybody had a model to ask, Drupal’s security team told the world in PSA-2014-003 that a Drupal 7 site not patched within seven hours of the Drupageddon announcement had to be treated as compromised. I have no such number for 2026, and I would not bet my sites on it being longer, so I am not going to brag about our six hours and forty-four minutes either.
The trouble is that a pinned release reaches a server with its next system pass, which on our hosted fleet means the small hours of Sunday, more than three days away, and on a self-hosted box with the automatic updates off means whenever its owner remembers. In October, on the day of The Ninety-Minute Recall, we built the forced recall for PHP and, I have to admit, only for PHP, as if the web server in front of it could never have a bad day of its own, so on Wednesday evening the floor had to learn about nginx as well.
I read the small print of both twice. Drupal’s advisory says, in its own text, “Discovered by Anthropic’s OSS Scanner”. Nginx’s change log thanks a company which sells AI-powered vulnerability discovery for four of the ten fixes, and the public record of the HTTP/2 memory corruption among them, scored 9.2, credits four separate reporters for the same bug, one noted as working with a GPT model.
What the defaults bought you is time, this week instead of this hour, and I would not waste it. The fixed releases, 10.6.22, 11.3.22 and 11.4.13, go into your own codebase by the usual route, which that page walks you through in an evening, and a door bricked up on our side is no reason to keep a known fault in your own code, the more so as the advisory’s smaller sibling, SA-CORE-2026-018, has no remedy but the update.
I know the face some of you make at this, I made it myself until this year and wrote it all down in Artificial Stupidity Revisited. But look at Wednesday once more. Several of the faults were found with machines, the records say so themselves, and the other side has the same machines. The attacks we write our IDS rules against already visibly use AI, as I wrote back then, and the diff which shows where to look is public the minute the release is.

Terminal window on a BOA test server on the stable lts tree, host name changed: the floor's log says at 21:30 UTC that Nginx 1.31.8 is below the security floor with a system pass due at 21:53, and at 21:55 that barracuda up-lts system was launched; the pass log shows Nginx 1.31.9 installed from prebuilt package at 21:57 and BARRACUDA upgrade completed; nginx -v reports nginx/1.31.9; and the floor's state file no longer exists

Back to the blog »
So what is left for you to do? On a site hosted with us, nginx needs nothing and Drupal needs the core update, this week, at your own pace. On your own BOA, nginx -v should say 1.31.9 by now, the floor’s log is /var/log/boa/sec-floor.log these days (it lost the php- in its name on Wednesday, for obvious reasons, and kept its history), and a box with _SKYNET_MODE=OFF mails you the minute it picked before it goes ahead.
I asked whether that still holds at about 18:55 UTC, a good hour after the advisory, and I did not want “we have always done it this way” for an answer, since that is a belief, and beliefs are what get servers hacked.
Nginx was the other kind of evening. Four of its ten fixes reach every BOA server, because every site here is served over HTTP/2 and HTTP/3 with TLS 1.3: three memory bugs in nginx’s HTTP/2 and HTTP/3 code, the kind which can crash a web server or worse, and a request body over TLS 1.3 which can keep a worker spinning for ever. Most of the others live in parts of nginx a BOA server does not use (the mail proxy, for one, is not even compiled in). For the four there is no switch to flip, since turning HTTP/2 off would slow every site down and still leave the TLS loop, so it had to be the new release, on every box, and the sooner the better.
For twelve years keeping that up meant one of us dropping everything else. It still does, only now the one of us has a pit crew. A pit stop takes seconds where a dealer takes a day with the same four wheels to change, because the crew rehearsed every move long before the race, and once the car is called in nobody in the pit lane stands around debating it.
This is one of our test servers, one which sits on the stable lts tree and runs nothing but the published code. It saw the floor at 21:30 UTC, half a minute after the mirrors had it, booked itself in for 21:53, started at 21:55 and had the new nginx two minutes later, from the prebuilt package, with the pass over before the hour.
Now the part some of you will not like. Between my “yes, build it”, soon after 19:00 UTC, and the floor on the stable trees at 21:30 there are two and a half hours, and in them the change was written, run through a sandbox of a hundred-odd checks on a real server, proven by a real forced pass which took a test box from 1.31.8 to 1.31.9 in six minutes, documented, and put through four rounds of review by separate AI reviewers, which threw one of the fixes out along the way (correctly!).
Fifty minutes later the answer was live on the security releases page, and a test server had already proved it: 84 requests against six sites, Drupal 7 to 11.4 and Backdrop, the installer in ten disguises and the rebuild, authorize and update scripts beside it, and 84 answers of 404, then the fixed 10.6.22 and 11.4.13 cores built on that server and a fresh site installed on each through the panel’s own task queue, to be sure that a release which tightens the installer does not break the one install a hosting system is supposed to do.
Here is Wednesday evening on BOA in short, for those of you with a train to catch. The Drupal bug needs a door which has been bricked up on every BOA site since 2012, and we spent the evening proving it on a test server instead of assuming it. For nginx no setting helps, only the new release, so 1.31.9 was the pinned release on our dev, lts and pro trees by 18:10 UTC and by 21:30 UTC a security floor stood under it, which means a BOA server caught on an older nginx books its own upgrade for some minute in the next thirty, whether its owner left our automatic updates on or not.
I did not do that alone, and no team of our size could. I did it with Claude Code, which read the advisories and the diffs, wrote the change, tested it on our test servers and argued with its own reviewers, while I asked the questions, made the decisions, corrected its course more than once and published every tree myself.
SA-CORE-2026-017 is about install.php, the script which sets a site up on its first day and should never run again, and which under specific conditions could be talked into reinstalling a live site, with the takeover you can imagine. The advisory names two things which blunt it, a secured installer and a web server which cannot write to settings.php or change its permissions. On BOA both are defaults on every site, and the installer is closed on the Ægir control panel too: Nginx sends the installer a 404, the same answer any PHP file gets here unless it is one of the handful a site really runs on, and settings.php is owned by the hosting account, with PHP working under another user which may read the file and nothing more.

Similar Posts