Kevin BeaumontSep 28, 2023

I wish I didn't auto delete my toots sometimes, as I predicted this about 6 months ago...

People are injecting malware responses into Microsoft's AI, so now when you ask it questions it is serving people malware downloads. https://www.bleepingcomputer.com/news/security/bing-chat-responses-infiltrated-by-ads-pushing-malware/

Bing Chat responses infiltrated by ads pushing malware

Malicious advertisements are now being injected into Microsoft's AI-powered Bing Chat responses, promoting fake download sites that distribute malware.

BleepingComputer
Show threadKevin BeaumontSep 28, 2023

From that thread (RIP), expect nation states, SEO spammers and more to be filling generative AI with crap to install malware, influence policy documents, research etc etc.

It's absolutely the next stage of enshittification (sorry, I mean increasing shareholder value) where everybody can pretend to shocked it happened in two years.

Show threadJohn FrancisSep 28, 2023
@GossiTheDog it sees the shit and is learning it. Makes sense.
Show threadLuna LacteaSep 29, 2023
@johnefrancis @GossiTheDog Doing exactly what it was made to do, learn. It can't unlearn either, unless they have backups of it to load.
Show threadJohn FrancisSep 29, 2023
@jackemled @GossiTheDog yeah, it's hard to see how anyone can claim to have undone the copyright violations they committed during training. So I guess they'll just have to license at whatever term the creator wants.
Show threadLuna LacteaSep 29, 2023

@johnefrancis @GossiTheDog "ooh it's not stored intact, the ai shreds it", so you admit it's not intelligent, & that it's just an idea blender?
"nooo it's original it makes new things, inspired by what it remembers", so it does involve stealing copyrighted material?

I wish they would at least make up their minds & be consistent with their bullshit.

Show threadJohn Francis
@jackemled @GossiTheDog I guess we'll find out when it starts spitting out mouse ears and Like Skywalkish and the AI companies eat Disney's lunch.
Sep 29, 2023 at 1:20amWeb
Show threadLuna LacteaSep 29, 2023
@johnefrancis @GossiTheDog I hope it's soon!
Show threadNatanael ⚠️Sep 29, 2023
@jackemled @johnefrancis @GossiTheDog somebody needs to develop a robust multistage ML poisoning scene. Single stage to make it respond with desired responses to targeted keywords already exists but seems too easy to revert, but if you can slowly plant the malicious responses over time and "activate" them by publishing another final set of seemingly irrelevant samples then you can plant all kinds of malicious responses that are VERY hard to delete from the model
Show threadLuna LacteaSep 29, 2023
@Natanael_L @johnefrancis @GossiTheDog YES