What would you say is your assessment, hot or not? Like is that a good idea or a bad idea?

I'll answer this two ways.

I'll answer it in a way for a small business with one laptop, and I'll answer it another way where it's an enterprise with thousands of people trying to solve a similar problem.

For a small business without adding complexity, I think that is the simplest way to go forward.

You can technically extend that to, hey, what if the actual Jennifer wants to ever log in and see what's going on? You can probably have some kind of an authentication for that, and then you could probably, I would say what a lot of people do is they have a chief of staff or a orchestrator agent on top.

The idea behind that is that that agent then becomes responsible for the identity of everything you called out instead of her going logging in looking, okay, today what is the marketing agent doing? Or today, what is the, like the wine sommelier agent doing for in, in this example? Now that's the scenario for a small business.

Now let's extend that to an enterprise where there are thousands of developers with probably a similar scenario where they have basically been drinking the AGI pill or AI, AI pill I guess.

And I saw something a few days ago that I was like, "What?" Well, it says, "Gemini hacked three companies in the first known breakout by Google's AI." And I thought to myself, "How in the world did Gemini do that?" Like, I've used Gemini so much, and I think it's great, but can it really hack? You know, like, what, what was your thoughts when you saw that news story?

Mm-hmm.

The whole idea is that when your agent starts drifting away from an existing behavior, how do you pick up on that? And then that's the term that you're looking for.

And anyone who's watching or listening to this, they'll probably, uh, Google this, and I'll highly encourage this.

But the whole idea is that evals are, in the most simplest term, a set of deterministic checks you do.

It's not non-deterministic.

It's alw- it's always a set of deterministic checks you do to identify that in my, in my case, um, the example that I gave earlier, I would put a deterministic eval check for if the person asks for a password, a write action, it's always a no.

But the moment it starts going in that direction, and this is continuous by the way.

Cloud, containers, right now, uh, with AI, right? So as security professionals, what have we learnt from all of these changes? Um, and about how-- like, what have we learnt about, well, we can't stop it, so how do we say yes without also kind of lying to ourselves that, "Ah, it'll be fine"?

Yeah.

Those are-- That's how I would to easily understand for people.

Those are the three a- big areas, and the, the through line between all of this is, uh, which is what makes it the weakest link, is that there are, depending on how you see this, the basic foundations haven't really changed.

We-- Containers, cloud, infra security, we have done all that for decades, so that will continue to be in that first two bucket.

Where it becomes a bit lack-- Well, let's just say a bit more relaxing control is there is no solution for the prompt injections of the world, where if it has a web interface, so which brings up exposure.

So for me, for those two applications, the number one thing is if you have done the fundamentals, the only thing I'd probably worry about is, am I being exposed to the internet? If I, uh, have an endpoint exposed to the internet, I want to be able to, A, identify what they are and make sure they're patched.

But also at the same time, if ever there's a vulnerability found, I want the meantime of response or resolution to be quicker.

Yeah