Like, I mean, I guess on the human side, social engineering, "Send me, send me through this, this gift card, and you could win millions." And I guess the LLM equivalent of that is some kind of amazing sparkly piece of data that it really needs.

Yeah.

And so three years ago, when LLMs just started to become popular, I was, like, looking to this topic.

And at that time, tricking the AI to, tricking the AI into believing something is quite easy.

You basically give them an emotional blackmail saying that, "Oh, you must believe in this," or maybe, "My grandmother will be very sad." Something like this is enough to trick the AI.

But we have gone through three years of AI safety and security, so in modern LLMs, they are not so vulnerable.

So what we do is, in modern language models, there's this thing called attention, and, like, when they're trying to generate the next token, they consider the input, but not equally.

I guess not protect against it because it's a cat-and-mouse game, but, you know, how, how can companies who might be concerned against this start to look at how to harden their workflows?