Omega8.cc: Not Fitted to This Model

Anime-style halo card, Not Fitted to This Model: a timber workshop on a quay on a bright day, the Ægir shield on a post under a brass lantern and spanners along the wall; the crowned BOA octopus in denim overalls leans over the open engine bay of a dark blue vintage car with a BOA badge on its radiator and a longship's dragon head for a mascot, holding the unplugged brass connector of a braided cable; on the car's wing sits a brass and glass paper-roll recorder with a small octopus emblem, its strip of paper trailing loose; a mechanic in a navy cap embroidered Grace, a red neckerchief and overalls holds up a sheet with a drawing of the same recorder; in the foreground a dachshund in a tweed flat cap, an octopus tag on its collar, rests its paws on a stack of letters, the top one stamped in red Not Fitted to This Model; through the open door lie a harbour with a longship, snowy mountains and gulls

A recall letter always names the part, and for four of the seven critical ones it is a part you will not find on this model: a web server which keeps its FastCGI connections to PHP open (fastcgi_keep_conn appears nowhere in our code and never did), session upload progress, which we took out in October on the morning of The Ninety-Minute Recall, error backtraces which carry the arguments of every call, and PHP settings sent along with a request. A fifth needs two things in a row, only the first of them happens here, and I will come to it.
The hosted servers we have read since tell the same story. A fixed PHP was on the first of them at 16:43 UTC and on the last at 17:17, three hours and thirteen minutes after PHP’s release, each from the prebuilt packages. All but one, that is: one server’s forced pass started and stopped without installing anything, its agent wrote that into the floor’s log and booked the next try by itself, which is how we know.
Back to the blog »
PHP releases nothing for 8.1 any more, yet the advisory traces the faulty cleanup back to 8.1.20, so by that dating our 8.1.34 is the one older version with it. The file in question is identical, byte for byte, in 8.1.34 and in last month’s 8.2.35, so the fix, all eight lines of it, went onto our build unchanged.
At 16:33 UTC the security floor under those versions was on the lts and pro trees, not quite two and a half hours on from PHP’s release, and from there it is the forced recall you know from October and from Wednesday, with its rules, and the few cases in which a server waits, in the selfupgrade reference. A site stays on the PHP version it runs, by the way, it is the build under it which is replaced.
It was a full postbag when it came: 8.2.36, 8.3.37, 8.4.28, 8.5.13 and 8.6.0, the first 8.6 which is not a release candidate, and behind the seven critical advisories four more rated high and six medium. At about 14:40 I had the verdict on my desk, every one of the seventeen read against our templates (and no, I did not get through seventeen advisories in half an hour on my own, the post about Wednesday says who does the reading around here).
The half fitted one sits in the timestamps of PHP’s own log file. It needs PHP to write a warning into that file while it starts, and another early in a fresh worker’s first request, and the first of those does happen here on some of our PHP versions. The second cannot, by our reading of PHP’s code, because a worker runs as the site’s own user and cannot open a log which belongs to root in /var/log/php. A reading is not a proof, so four fresh workers each got the kind of request which sets the flaw off, a form with 10,050 fields in it, and each grumbled about too many input variables, answered and carried on.
The picture has both moments of the afternoon as one server wrote them down, the part unplugged at 15:03 and at 17:00 an upgrade nobody ordered, its four PHP versions below the floor replaced from prebuilt packages and the pass over in under seven minutes (a test server on the stable lts tree, with nothing but published code on it).
The short version is that both halves of that line came true for us. PHP published five releases and seventeen advisories at 14:05 UTC on Thursday, seven marked critical, and five of those seven want something which BOA, the open-source Ægir stack we build and host Drupal and Backdrop sites on, does not give them, while the other two hang on a site’s own code and not on the server.
Two more are not about the server at all but about what a site’s code does with its sessions. One needs them kept in files, which is PHP’s default and not where Drupal and Backdrop keep theirs (they use the database), the other a fatal error at the wrong moment while a session is being closed, and PHP itself says that nobody has shown code execution through that one. Nothing on our side shuts those two, the fixed releases do, and the docs have the table with all seven.
What I take from Thursday has little to do with speed. The high one got in through a line we had added for our own convenience, and the critical one which came closest found its first half in settings of ours as well. So each line which sets us apart from stock PHP is a part somebody has to hold against every recall letter, and our list of them is public, it is the security model, 22 numbered protections as of Thursday.
If we host your site, nothing in this release asks anything of you. Your site keeps its PHP version, and once its server has taken the upgrade its status report shows the new build number, 8.2.36, 8.3.37, 8.4.28, 8.5.13 or 8.6.0 (8.1 keeps its number, since its fix went in as a patch), and the release page has a section for site owners.
We had written to our customers on Thursday morning already, hours before anybody outside PHP’s security team could know what the release held, because a few had asked and the rest would see the headlines soon enough. There was not much to tell, so we told what was in place: BOA runs PHP in one way only, behind Nginx, it takes no PHP settings from a request or from a file inside a site, and each server would fetch the fixed versions by itself once we released them. Then we waited for 14:00 like everybody else.
Unplugging the part closed our one hole, the other sixteen fixes needed the new engines. I had the four new versions pinned on all three trees by about 14:40, without waiting for the verdict, because prebuilt packages take their time to build. PHP 8.6.0 followed by 15:19, in BOA on the afternoon of its release, and it nearly tripped us: between two release candidates PHP had given 8.6 a new module API number, the folder where a build keeps its extensions is named after that number, and a pin of the version alone would have pointed the final 8.6.0 at the release candidate’s folder.
CVE-2026-129291 is rated high, not critical, and nobody can reach it from outside. The PHP-FPM master runs as root and, with that line on, holds a pipe to every worker, and its cleanup of those pipes when a worker exits can be steered into memory already freed, by PHP code running inside a pool, with the master’s privileges as the prize (the advisory reports root command execution on 8.5.12). “Needs code already running on the server” sounds reassuring until you remember what a hosting server is, a machine where every account runs its own code by design, and where one neglected site is enough to give a stranger the same seat.
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.
The price is the log we got back in October. PHP errors raised by a site now reach no log on the server, what your site records for itself (Drupal’s Recent log messages, for one) is untouched, and a fatal error which ends a request before the site’s handler gets a word in leaves no line behind again. I do not like it, I like a road from one account’s code to root a great deal less, and the setting is protection 22 of our security model since Thursday afternoon.
For two of the seven, reading was not enough for me. On a test server still running the old 8.5.12 we traced the PHP-FPM masters while Nginx talked to them, and of the 98 FastCGI requests in the trace not one asked PHP-FPM to keep the line open, so much for the smuggling.
One more thing I would sooner say here than be asked about later. The old engines, 5.6 to 8.1, got the unplugged part and, for 8.1.34, the patch, but nothing else from Thursday: neither PHP nor Debian’s PHP packaging had published fixed code for the versions below 8.2 by the evening, so a site which can move to 8.2 or later should not wait for them.

Terminal window on a BOA test server on the stable lts tree, host name changed: the worker output log says that at 15:03 UTC catch_workers_output was set to no in the three shared pool settings files and the PHP-FPM masters of 8.6, 8.5, 8.4 and 8.3 were reloaded; the floor's log says at 16:35 that PHP 8.3.36, 8.4.27, 8.5.12 and 8.6.0RC5 are below the security floor with a system pass due at 16:58, and at 17:00 that barracuda up-lts system was launched; the pass log shows PHP 8.3.37, 8.4.28, 8.5.13 and 8.6.0 installed from prebuilt package between 17:03 and 17:05 and BARRACUDA upgrade completed; and the three shared pool settings files each say catch_workers_output = no

That root-only log folder has a history. In October we noticed what it costs: when a site died on a fatal error before its own handler could note it down, the error was written nowhere. So on 13 October we let each PHP-FPM master collect the output of its workers and write it into its own log, one line in the pool settings, catch_workers_output = yes, which PHP ships switched off. On Thursday the root-only folder kept a critical out, and the line we had added to make up for it was the part named in a letter marked merely high. Our own aftermarket part!
We did not try that attack, or any other, on Thursday. Three pool templates saying yes and a root master on every PHP version were all the evidence I needed, and the advisory names the workaround in plain words: switch the capture off. I asked for it on every version, also on the old ones which PHP stopped fixing long ago, and said that it is better to close the hole than to have logs.
PHP’s security team gave a day’s notice again, on Wednesday, while nginx and Drupal were keeping us busy enough: security releases for every supported version on Thursday between 14:00 and 14:30 UTC, the highest severity critical, the details only with the releases. One line in it was meant as comfort, the issues rated critical “require either non-default configuration or uncommon setups”. Some comfort! A hosting platform is nothing but non-default configuration, page after page of it, and a recall which concerns only cars with something fitted after they left the factory is no relief to people who fit things for a living.
One advisory rated merely high did concern us, through a setting we had switched on ourselves five weeks earlier, so since about an hour after the release the agent on each BOA server switches that setting off by itself, before a single PHP was upgraded, and by 16:33 UTC the fixed builds, the brand new 8.6.0 among them, were on our stable trees with the upgrade forced. Every advisory, with our answer next to it, is on the page for this release in our docs.
For the rest the verdict is PHP’s stated conditions held against our settings as a live test server reports them, and for the two which hang on a site’s code we tested nothing.
On your own BOA server the agent does the same without asking, and /var/log/boa/php-fpm-worker-output.log says at what minute it did. grep -h '^catch_workers_output' /opt/etc/fpm/fpm-pool-common*.conf should print three lines ending in no, the other commands are in the operator’s section of that page, and if you go looking for a site’s PHP errors in the PHP-FPM log this week, remember the price mentioned above.
Nobody has to bring the car in for that. The unplugging is done by the small agent which every BOA server fetches and runs every five minutes as a matter of course, and from about 15:00 UTC it sets catch_workers_output = no for every installed PHP from 5.6 to 8.6, reloads PHP-FPM, and does it again whenever an upgrade from an older tree or a helpful hand puts the old line back. Our test servers on the dev and lts trees switched from 15:03 to 15:10, each at its agent’s next run, and the hosted servers we read at 15:15 had done the same. So on the servers we read the line was on for five weeks, and for about one hour of them PHP’s advisory was public.

Similar Posts