Not completely convinced that the "Oh, our AI broke loose on its own and attacked another system" story tells the whole story.
4
Replies
·
0
Boosts
I have used my super-duper AI hacking skills to find out what the containment was, and it's actually just a single instruction to the AI:
"plx don't hack other companies' data and steal it for us UwU, even though we could make lots of money and money would make us happy *wink wink, nudge nudge*"
Hmm, seems foolproof to me! Must be super-AI!
(yes, all the things in this post are made up, but at least it was made up by a human, not an LLM)
@lauren If the crappy AI managed to break out of its containment, doesn't that rather demonstrate what a shitty mess their (surely vibecoded) security/backend stuff is? It's very obvious this "warning" was intended as a viral marketing attempt, but as a technician it only makes me want to touch their software even less.
@lauren
1. Like the blackmail stories, when you look at it, they had to train the system with specific blackmail to make it do that.
2. Negligent release of an automated system that cracks other os a computer misuse offence well established. The first worm was released on the network accodentally so it's not as if compsec reseaechers wouldn't know to properly sandbox any white hat system.
@lauren I'm sure there's no reason for the marketing team at OpenAI to over hype things after Anthropic released a model so dangerous it got banned 🤔
