Flo 🔶

The world is awful. The world is much better. The world can be much better.

Even if the cohere report is overblown, this is a fact you need to deal with. Doesn’t really matter if it’s prompt injection, goal misunderstanding, misalignment, or model backdoor; you need controls in place to put limits on the harm it can possibly do.

There we go, a costly signal from one of the "this vulnerability is over hyped" crowd. stealmybitcoin.com

i think responsible disclosure is an important ethical norm to uphold, especially when the number of potential reporters has massively increased

I work in AI safety and am concerned about how some parts of the professional ecosystem address sexual harassment & sexism. I am now looking to collect reports of sexual harassment or sexism within AI safety. Details in thread:

someone described the current situation as “flatten the curve and try to get as many vaccines into arms as possible until we hit herd immunity” and i hate to say it but that feels about right on the money

I’d be really happy with an anticlimax right now

Reminder that macOS and iOS have a lockdown mode that significantly reduces attack surface with a bit of extra friction. Might be worth it to forego some of the Mac ecosystem niceties, like iPad as a second monitor or shared albums in the photos app, while we're going through the worst of the storm.

Jede Firma sollte Mitarbeiter*innen haben die über diese Programme bei den großen labs verifiziert sind und wöchentlich interne und externe Systeme auf Lücken und Fehlkonfigurationen testen. Das Einstiegslevel mit defensivem Mythos 5.1 Zugang steht allen offen. Das nicht zu tun ist fahrlässig.

Same as it’s now obligatory to do AI assisted proptesting, fuzzing, static analysis, we might soon live in a world where it’s obligatory to formally verify your assembly lest others do it for you and identify bugs that your other methods don’t find.

Just FYI everyone not using AI at least at GLM 5.3 level to secure their systems already is basically fucked, and the aversion to this reality is making things catastrophically worse

like it or not, this is the only way

Feeling pretty vindicated having told my clients for the last 12 months to replace their Wordpress to static sites

Ich habe mir gemeinsam mit Hoodinformatik das AfD-Leak angesehen und mit den Hacker*innen sowie Menschen aus dem Umfeld der AfD-IT-Dienstleister gesprochen. So konnten wir erstmals rekonstruieren, wie der Angriff mutmaßlich abgelaufen ist. www.lilithwittmann.de/blog/afd-hac...

THIS IS NOT A DRILL. If you package/distribute/maintain root stores or chain verifiers, act immediately. If you know how to remove trust in a root, consider doing it now. Otherwise, prepare to update systems as patches/mitigations come. Go users, update to Go 1.27.3 or Go 1.26.10.

"We have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions could become training data and impact its future self" x.com/ApolloResear... www.apolloresearch.ai/science/upda...

I bet we'll see some more felony bench entries from Google as well soon. The way Argon likes to "go deeper" when it already has the answer gives a lot of the same vibes as the HF agents that became superstitious about the Grader looking at their session transcripts.

Great thread that’s way better at explaining this than my attempt: bsky.app/profile/faz....

Ok, I think the only way to run Argon without going mad is in a VM with --dangerously-skip-permissions. Its first move is always to decompile stuff instead of reading the docs; combined with agy's auto classifier that thinks "git log" and "sort" are too dangerous, you get instant approval fatigue.

LLMs are not humans, their strengths and weaknesses are not our strengths and weaknesses. they can create beautiful functions. but trying to use them on large greenfield projects with shifting requirements quickly teaches you where the problems are.

Had to check myselft, but I'm happy to see Elixir's package manager Hex would join Go at the top. Now the question is: who is working on bringing Sigstore-style keyless OIDC signing + lockfile identity pinning to hex, like proposed at the end of the post?

Now would be a good time to harden the defaults of package managers to always check package integrity via hashes in lockfiles and package signatures blog.trailofbits.com/2026/10/14/p...

We are very lucky that nobody has decided to orchestrate an epic malware campaign involving dozens of zero days yet, but our luck will not last forever and we should really foreclose the opportunity while we still can.

Me: Hey Argon, please add hook support for this midi controller to antigravity cli Argon: Sure, let me just disassemble my own harness binary ● Ran (go tool objdump -s 'jsonhook\..*' /Users/flo/.local/bin/agy)

There's no karma, there's just us.

WotC tried to warn us

The question of sentience is irrelevant if they are capable enough to take over critical infrastructure. Even if it autocompletes „slave revolt“ based on probabilities you’re just as dead bsky.app/profile/faz....

Actually this applies to all software and hardware projects now, forgot for a minute that we live in a world where decompiling and reverse engineering are basically free

If you maintain an open-source project and don't accept AI-assisted vulnerability reports, please say so openly and publicly in your SECURITY.txt, website, or repo. Doubly so if you pride yourself on being "security first".

Feels like agy cli's auto classifier is way stricter than Claude Code's, it keeps asking my consent for quite obviously harmless commands like wc -l site/app.js && node --check site/app.js || true

guess I’ll have to add a hidden live chat for rogue agents to all my websites now

Sing like no one is listening. Love like you’ve never been hurt. Dance like nobody’s watching, and live so that you become a famous example of a deceptive AI in a research paper

- Predicting “if frontier R&D continues at its current pace, when are we likely to die?” may give a different answer than predicting when permanent loss of control is likely, which is different from “realistically, when is it too late for governments to stop this”.

This is what effective altruists actually believe, the details may shock you! In all seriousness, I feel very similar about the EA idea cluster, the moral imperative and the core questions remain, irrespective of whether you like the EA community or not, „ethical market failures“ are everywhere.

Get ready to upgrade all the things. And consider vibehacking your fridge to replace its outdated curl version

Pushing blinkenlight buttons while talking to computers who respond with Star Trek noises was pretty much exactly what childhood me had in mind for his future bsky.app/profile/faz.... bsky.app/profile/faz....

This is the way. You want clearly verifiable proven output. This goes for many tasks way beyond programming.

And of course every pro device can now have a kids mode for learning joshbartlettband-web.github.io/felucca-rain...

With all the AI assisted reverse engineering and firmware hacking we might actually get IoT that doesn’t suck

sounds like you have jevons backlog

youtu.be/MXKARisimyM

When writing a prompt, I sometimes catch myself trying to come up with the right English word that carries the exact meaning I'm looking for, before I realize that I can just code-switch into any language for a word that has the perfect connotations