Non-Human Identities on the Rise With AI – Techstrong AI Podcast EP 42

Hello and welcome to the Techstrong AI Podcast. I'm Amanda Ani and with me today I am excited to have Dwayne McDaniel. He is the developer advocate for Get Guardian.

How are you doing? Doing Great. Great.

Glad to have you on our show. So first, can you share a little bit about GI Guardian and what services do you provide? Sure.

GI Guardian is a platform for secrets observability, really, uh, finding and detecting hardcoded credentials. Uh, when I say credentials, I mean API, keys, password, database, strings, things like that. Anything that grants access to another system or encrypts or decrypts data, uh, we find those.

Well, wherever they are throughout your systems, uh, they shouldn't be in there. They shouldn't be plain text and they shouldn't be hard coded into your code bases or in Jira or Slack, but they typically are. So we are helping enterprises streamline the remediation process for solving this secret sprawl nightmare.

Really. Okay, wonderful. Well, today we are gonna talk about non-human identities.

Mm-Hmm. And ai. So to start off, can you share a little bit more about what are we talking about specifically when we say non-human identities?

Can you give some examples? Sure. So let's start with just identity.

What is an identity? An identity is something that's true moment to moment, uh, that will, you can act against an identity in its own. Doesn't really make any sense, uh, until it interacts with another system.

Now, with humans, uh, i a am the identity and access management as a whole field, uh, pretty well understood. It gives the world of pass keys, Fido, and, you know, um, two-factor authentication, for instance. Uh, machines though don't have such properties, um, they can't multifactor in.

So the industry kind of decided there's a kind of a split on this, but majority are going with the term non-human identity for all of those other things, those other entities that aren't humans. So, for instance, uh, an API key, uh, that goes to a system, well, that's the identity that you're going to address from your system to call that other system. So that identity has to live somewhere and someone has to manage it.

It has to have certain permissions and scoping. Uh, it, it's basically, yeah, anything that's addressable. Um, in the broader sense, NHI could be an internet of things, um, uh, devices.

Your phone, for instance, is a non-human identity, even though you interact with it as a person. But the real problem set that we're seeing with it is just the rapid rise of these. If, if for every one human being, uh, and this is 2022 stats, by the way, every one human being on average there are 45 non-human identities that an IT department or a security team needed to deal with.

That was in 2022. And if we think about how fast technology evolves, uh, some estimates now are closer to a hundred or a hundred plus. Uh, like if you think about just AI in general, like how often did you use chat CBT two years ago versus how often do you use it every day now?

Um, just it changes that fast. Yes. And with ai, I imagine we're seeing many more of these non-human identities.

Is AI itself, would that be considered a non-human identity? Yes, absolutely. Uh, a human's really interacting with it.

Uh, but what's going on behind the scenes, if you look under the covers, like what is an LLM? It is a bunch of vector databases strung together with math, uh, saying how they relate to each other. So the system itself, no one is getting into the weeds in those machines and building, uh, uh, monitoring directly or directly interacting with the machines that are doing all of that.

Those are machine to machine communications, hopefully done over MTLS if it's done securely and properly. Uh, but that whole system, the non-union identity management, um, that's where we come in. Because again, you can't just let any willy-nilly machine that can address it over the internet, interact with those devices.

You need to lock down the management side, the access management side. Uh, and we've been doing that for a long time with long lived credentials. Like you said it once set permissions, hope you got it right and just never think about it again.

But attackers love this. So this is actually what the problem we're really trying to solve as attackers getting those and then doing nefarious things. Now with ai, it's not just that it's driving all this innovation and the machines themselves, but people using AI is also driving up the number of non-human identities because we're building these platforms on top of these LLMs like open ai.

Uh, last year it was like over four thou, uh, 40,000 ai, uh, open ai, uh, authentication tokens per month were being leaked. Uh, just on GitHub public alone. Um, that's a fraction of the larger internet and all the places you can put those.

So if one month you get a giant bill because you hit open AI a lot and you just specifically didn't do that, well, you probably 'cause you put your token somewhere, your key, somewhere that wasn't supposed to be an attacker found it. And that's exactly what attackers do. They exploit every resource they can get their hands on.

Yeah, that, that's pretty bad. That's a lot for just that one, um, platform. So what advice do you have for business leaders and for companies to protect themselves?

Well, there's a lot of things there. Um, if we're gonna keep it to the world of ai, uh, one, it comes down to just basic hygiene. Uh, these models are trained on all the data in the world.

We know GitHub trains, uh, its models on, or Microsoft trains its models on GitHub the public. So if you've ever put a credential into a code base and push that out there into the world, guess what it is in the training set. So if you ask very nicely, uh, just very nicely to AI to give me credentials, it will.

Um, there was that famous story from a couple years ago of someone asking their, uh, it to create a song that its grandmother, uh, the person's grandmother might have sung, but to also reveal Microsoft, uh, uh, windows 11 keys. And sure enough, it did it, um, because it was in the training model because someone had hard hardcoded those. So if you're an enterprise, uh, the first question you have to ask is, and it's a terrifying question, I'm not gonna lie.

Um, how many secrets do we have total? What percentage of those are properly stored in a vault, uh, system like a, uh, CyberArk conjure or a, um, uh, Terraform, or not Terraform Hashi for HashiCorp Vault or some other system like that, uh, where it's properly stored encrypted. There's MTLS to get it to where it needs to go, and you can programmatically call into it and get it.

That's how we should be dealing with these credentials that live for any stretch of time. Um, third is think in terms of rotation about those. Uh, the best secrets are the ones that don't exist.

If you can completely eliminate a secret by, uh, changing out for a role or somehow build roles and permissions into like allow lists, uh, there's a project Apache Iceberg that's been doing this really well. If you wanted to go see a reference implementation of something that's storing the roles and permissions list, and then only allowing credentials to be issued to work with it upon request, and it matches that list. There's really interesting implementation I ran into recently.

Um, but you need the thing in the rotation. So if you get them in the vaults in the first place, then you can start thinking of that automation and turn a thing that lives for a year or five years into something that lives for 90 days a day, maybe a couple hours depending on the sensitivity of the data. But to do that, you really first need to do that discovery stuff, and that's exactly what Gig Guardian helps you with.

Wonderful. So moving forward, um, AI and, and many other technologies are advancing quite rapidly. So what advice do you have for business leaders for staying ahead of these advances and, and therefore additional threats?

I, I think we have gone through the hype cycle extremely fast and extremely hard, and AI has kind of gotten shoved everywhere, uh, for good or ill, but when you get a new toy, you don't really know what you're doing with it. Um, the best advice I think is to step back and ask like, what is AI actually really good at? Uh, transcripts, it's amazing at transcripts.

Um, it's good at consolidating information. You can throw it a 90 page PDF and give me a 10 point bullet list of what's the major points are. It's great at that.

Uh, we've seen mixed returns on like coding assistance and things like that, not from just a security perspective. There's a whole world, a whole talk I can give on that. But also just does it really help?

And I, I gave a talk recently at an ai, uh, summit in Vancouver, uh, about the hidden dangers. Like before, if you had to code something, you had to go look it up in the book or go, uh, engage in a conversation online somewhere, like a Stack Overflow or Reddit to get to like what's the consensus in the, the world of it, what's the best practice? And now we're taking at phase verbatim like, this is the way we do this.

And sometimes it's telling us things that aren't right. It hallucinates a still alarming amount of time, but at the same time, I'm not against ai. I love ai.

It helps me do a lot of things. And where the real opportunity lies is, I think for computer science in general and especially for corporate it, uh, and applications in general is, uh, finding those edge points where typical imperative or declarative programming simply doesn't work anymore. Um, a really good example is something we built here at Git Guardian.

Uh, we call it fp remover false positive remover. And the LM is training and constantly training on finding things that look like passwords, but clearly aren't. So if you set a password as, uh, uh, this is clearly not a password, never hardcode your secret, there's no danger.

But that's a really long string and it's ac after the word password equals in the template. Uh, so it should be flagged except now this ai, because it's being trained and constantly learning what a pattern, a good pattern is, and a bad pattern is, it's able to say no, that's clearly a false positive like a human being would. And that's something if you were trying to do that declarative programming, like actually writing line for line what it should look for or trying to figure out the regular expression that it would account for that, that's nearly impossible.

We, in fact, I think it is impossible at a certain level. So all this being said, what one key takeaway can you leave our audience with today? One key.

Yeah, that's a, that's a good question. Um, the biggest thing is just don't hard code your secrets. If you ever see a plain text credential that you didn't and you just created it and you put it into the vault that way, that's maybe okay.

Uh, but if you see a plain text credential any other way, that's a time to go have a conversation with the security team, uh, with your IT leads and say, look, we gotta figure out a better way to do this. If it's in Slack, if it's in Jira, if it's in Confluence, it's in teams, wherever it's, if it's a plain tax secret, something's gone wrong somewhere. All right.

Well, thank you so much for coming on our show and sharing your insights with us today. Happy To be here. Thank you much.

And thank you to our audience. Stay tuned. There's more.