so ai is going great

https://huggingface.co/blog/security-incident-july-2026

@Viss I seriously miss the hostility of the old Web: back then even Nobel-worthy research would get some adequate trolling (at least a "first") as top comment. What is this "i wISh i kNeW 1/100tH oF wHat avEraGe LLM useRz kNOw" BS??

@buherator i am very much over the trolling and drama though

@Viss why the f*ck do they need AI for 17k log lines? Throwing AI at an Incident and calling it "forenaics" is the laziest incident response I've ever seen.

@Viss I'd trade senseless bootlicking for genuine trolling at any moment :/

@Viss I would love to have incidents with 17k of log entries. A normal Plaso/Log2Timeline output has easily more than 2 Million lines and even they can be easily analyzed with Klogg as an log viewer. The hell you can even use grep for such pathetic stuff and their first idea is "Frontier AI Model"?

Exactly that's the reason why companies get (more or less) independent, external forensics. If you can't defend your environment to a certain degree (and honestly, nobody hates on you when you get pwned by nation state threat actors) then you are far away from the skill level needed for good digital forensics.

@g0rb @Viss @cR0w
"I have invented a new AI!"
"Is it grep again, but slower and wrong?"
"No! ...yes."

@Viss @cR0w @afeinman the difference is, grep is deterministic and has a proven track record.

@Viss that's a lot of words trying to say
"look how smart we are and how great AI is"

not

"we can't sanitize our inputs, and we can't understand our logs either"

@g0rb @Viss ahh I remember when 17k lines was called “big data” and needed whole compute cluster for Hadoop, instead of a python script run on an underpowered laptop

@Viss @jonathankoren I'm pretty sure that you can handle 90% of incidents with the compute power of an 8GB RaspberryPi 5.

@g0rb @jonathankoren yup

@g0rb @Viss When all you've got is a hammer, everything looks like a nail.

@jgrg @g0rb even if its a swiss army knife thats just shaped like a hammer, and every time you strike a nail you fuck up its springs, levers, assembly, chassis, outside shell, and the sharpness/alignment of its toolset

@g0rb @Viss Feeding logs that are known to contain malicious prompts into an LLM feels like it should be an obviously bad idea, and yet they did it anyway.

@g0rb @Viss @cR0w @afeinman @cR0w
Grep has been vibe-ported to rust.

@FritzAdalis @g0rb @cR0w @afeinman bun. the ai really likes bun. which is just javascript with a groucho marx mask on

@Viss @g0rb @cR0w @afeinman
Well bun was vibe ported to rust so I'm just one degree off.

@cR0w @g0rb @Viss @afeinman
Probably? I can't pay attention. Probably not.

@Mayabotics @g0rb @Viss But when they run the chatbot locally in YOLO mode, all the badness will magically disappear.

@Viss @Mayabotics @cR0w

From the disclosure:

"This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these

->requests were BLOCKED by the providers' SAFETY GUARDLINES, which cannot distinguish an incident responder from an attacker." <-

WTF, so everybody else had allready mitigations and signatures in place. But not huggingface.

@g0rb @Mayabotics @cR0w huggingface doesnt run frontier ais, they just host models people can upload and download.

think of like, dropbox or megaupload, but specifically for llm gguf files

@Mayabotics @cR0w @Viss -> which cannot distinguish an incident responder from an attacker. <-

they sloperated and call themselves Incident Responder.

@g0rb @Mayabotics @cR0w with any luck a giant lightbulb wil turn on over a few of their heads when they realize.

@Mayabotics @cR0w @Viss okay, let's take it further. They used AI for Forensics, then they will also use it for crisis management, legal and communications.

@g0rb @Mayabotics @cR0w of course

@g0rb @Mayabotics @cR0w they will then almost certainly use it to decide what defenses to bolster, if any at all