shits all calvinball anyway, gotta do something different than whats been done to death if you want to find unique bugs.

if you have to pay that for several runs, across a lot of files.

eventually you are gonna probably have to start prompting with "look at the use of STRUCT in SUBSYSTEM" and eventually even more focused, the broad prompts are gonna run out of juice soon.

mythos is less interesting than a lot of people are making out, but the vibe i got from people actually using llms for bug discovery is that different models are finding stuff at least for now.

long term, oh god $50 to review a few files properly is gonna get costly fast if the rate of bugs drops

what i kinda want to actually see is a proper breakdown of the bugs being found, across different projects. like i imagine they are good and finding specific things, but how much reasoning across functions, length of hardcoded strings, etc things affect discovery

kinda curious how long this will go on for. i assume they'll tamper out in a few months or so, amount of collisions is apparently very high and the sort of obvious attack surface to prompt against is kinda limited.

chromereleases.googleblog.com/2026/04/stab...

that is brutal number of cves in one release.

Continuing on the swiss tradition of crypto ag!

(also props for getting ToBs to comment, nice to see an reputable firm comment in articles like this)

CODE RED. THE JERRIFICATION HAPPENED.

RETURN TO THE BUNKER NOW

btw when discussing grub, please note a large number of filesystems are disabled upstream for secboot anyway (implemented by a check for lockdown mode in the modules init func).

bit bizarre to see people cite the hfs CVEs when that hasn't been enabled in signed images for several years.

UKIs with systemd-boot, and having is some crypto keys locked behind measure boot (so you get a direct indication that the software changed or not by it not autobooting) is probably still the best setup.

the limited configuration format of systemd-boot is what makes it the strongest.

you can kinda do a hacky templating system with envblks and pgp signed configs, though its limited and probably not suitable for booting a lot of systems.

(and there is a bit of a thing with how you can change the behavior of commands and stuff via variables, though i couldn't do much with it)

grubs main attack surface is still the very expressive config file, which tbh i'd like to see the distros go down a route to lock that down a bit more if we are gonna actually try for secboot being meaningful.

hard to exploit the FS bugs if you don't get much control on the memory layout.

* filesystem parsing / disk encryption is way better left to the initrd, though we need to get that authenticated, but seems ubuntu is heading in the UKI direction at least. does at least get measured into pcr 9.

* image parsing has no place in a bootloader, though it has been fuzzed pretty hard by a few people but still some recent bugs.
* i never looked at raid/lvm stuff, and partitions only for stack clashing (not much attack surface in grubs implementation of partition schemes)

so its hard to be confident in it. does contain an ok spray primitive in `zfskey -p` though there are better ways.

* xfs/btrfs weren't too big of an issue from memory, code size seems pretty small. but not really used this early in boot so not that big issue of an issue to remove.

* the zfs code is several thousand lines across multiple files, and is a nightmare to fuzz. pretty sure i tried at one point, never hit anything but might have not patched all the checksums and stuff. hard to get suitable sample filesystems, has a large minimum volume size so perf sucks

discourse.ubuntu.com/t/streamlini...

probably a good move, amount of setups this causes issues is actually tiny (people misunderstood it quite heavily)

still signed grub across all the distros with different modules means all you really have to do is replace two files for generic secboot bypasses

Got it to be significantly less bad, though I still need to remove my badlink thing but that should be doable.

most of the runtime now is linked properly with a normal linker script, no longer having scripts to do macro defs when building the objects.

defining things at link time is much nicer.

github.com/bahorn/skp/i...

geniunely what the fuck was i thinking back in 2024

think today might finally be the day I do the linker script fixes I've been needing to do for ages.

1v1 dual to the death with me vs the linker to finally make this part of the project not the most disgusting piece of code i've ever written.

people be normal about systemd challenge 2026 (IMPOSSIBLE)

read the news today

This animal is so special…🥺 #sengi #cuteanimals #shorts #zookeeper YouTube video by Hertfordshire Zoo

www.youtube.com/shorts/RFOLq...

seriously you've got to see these guys

the algorithm is giving me opossum and sengi content, so i think i got that nailed down

do you at least get a free barrel with your oil now?

i think we need to start gatekeeping linux again

interesting times ARE CANCELLED. i no longer want to live in interesting times.

skate story is a good game, do recommend, beat it over the last few days.

top tier soundtrack, reason i found out about it was listening to a ton of blood cultures in recent weeks.

I don't even hate LLMs, but randomly PRing a full refactor and piling on the emojis....