S7 E11: Talking enterprise AI vulnerabilities, jailbreaks, and active threat intelligence with Amy Chang

I love that.

I love that.

And of course, you know, cybersecurity is just a broad field.

Is there a difference? I mean, we're in the AI era now.

Is there a difference between just cybersecurity and AI security or AI cybersecurity?

There are a lot of commonalities, and then there are a lot of differences.

So cybersecurity is kind of like the confidentiality, integrity, availability of our digital systems.

Beyond that, I think that those concepts still broadly apply to AI, but I think the way that they are built and the way that we use them today makes them materially different in how we think about it because they are non-deterministic, next token predictors.

They don't think for themselves, but they are also...

By being non-deterministic, it also means that you can't necessarily predict their behavior.

That's

right.

And so the way that we have to think about security from that vantage point then is more of like how then do we leverage natural language as an additional element to consider when we're thinking about securing AI systems.

Beyond that, too, I think there are a lot of elements in the AI supply chain that are also unique to the way that the models get built.

You and I, or an average person, would just kind of interface with the model endpoint themselves or the model themselves, but don't take into consideration all the data and the training and all the models that went into the training of the models that you and I use day to day.

Right.

Right.

Wow.

I mean, there's so much to consider.

I mean, you know, you mentioned non-deterministic.

I just talked to somebody on another podcast about that.

I didn't realize just the extent to which things are really unpredictable.

So it's a great point.

Exactly.

Yeah.