Podcast Transcript

Welcome to the RSAC cyber at the top podcast where security leaders across industries share the strategies, hard lessons, and real experiences shaping modern cybersecurity and AI. Insights are grounded in community and built for every level of profession.

Hello, listeners. Welcome to this edition of the RSAC cyber at the top podcast series. Thank you for tuning in. I'm Tatyana Sanchez.

And I'm Kacy Zurkus.

And we are your RSAC podcast hosts. Kacy, what are we going to discuss today?

Well, Tatyana, as you can expect, there are new phenomenons evolving left and right every day. And today, we're talking about a new phenomenon in software. Our guests have a name for it, app generators, in case you haven't heard of that. These are platforms that are built on the same back end as a service platform, and they can leave applications vulnerable.

A single weak spot could open the door to every one of those vulnerabilities. So today, we're excited to be joined by Ari Marzuk and Roy Nassimi as they walk us through a real world technique that shows how attackers plan and run campaigns designed to hit many victims in just one go.

We'll also look at how certificate transparency can turn just one single misconfiguration into a target list that keeps updating itself with more than 40,000 potential leads. Lots of crazy stuff.

Definitely ready to dive in. But before we get started, we wanna remind our listeners that here at RSAC, we host podcast twice a month, and we encourage you to subscribe, rate, and review us on your preferred podcast app so that you can be notified when new tracks are posted. And now we would like to ask our guests to take a quick moment to introduce themselves before we get started. Ari, let's start with you.

It's great to be here. I'm Ari Marozuk, also known as Makarita. I've been doing cybersecurity mostly on the offensive side for the past ten years or so. I'm currently doing AI security for Salesforce and have previously worked for Microsoft and NSO Group. I'm also the person behind IDE Zoster, the vulnerability class that previously affected all IDEs with dozens of CVEs. However, we're here to talk about something new. But how about Roy introduce himself first?

Thanks, guys. So much fun to be here. Thank you for hosting us. My name is Owen Esumi. I'm a principal security researcher at Orca Security.

I'm in this business of hacking, finding vulnerabilities, and whatnot for the past twelve years. I've started my journey in the intelligence corpse of the IDF. And ever since, I worked for Orca Security for the past four years now, protecting our clients, finding vulnerabilities, and also contributing to the security community.

Great. Thank you both for joining us today.

So, Roy, to get things started, can you explain to our listeners what app generator platforms are and what makes them so unique from a security perspective?

Yeah. Let's go right into it. So we actually named it a few different names. It started as AI app builders, and then we we changed the name to app generators because it sounds more tricky, I guess.

But to to summarize, these are platforms that you may know and heard about, like Lovable, Base 44, Replit, v zero, and others that are less known probably. These platforms are pretty unique because they give everyone the opportunity to actually build a web app end to end. So we know everyone can use cloud these days, but not everyone are tech savvy. So if someone who is not really into tech and understands security and web development, it's pretty difficult to actually know what to ask for cloud.

So these platforms, they just give and hand everything to a non tech personnel. So we can just say a few words and have his online business ready, hosted, and ready to service customers.

From a security perspective, this is really interesting because that's the first time ever in our time that people who know nothing about tech ship products faster than anyone else without core review, without security development cycle reviews. And it's it's specifically interesting when people from within enterprises who are not techies actually use it. So imagine you have people from sales or marketing actually building apps, integrating these apps with sensitive data, and shipping them. No code review, no SDLC, nothing. So it made these platforms really interesting to actually identify and see where they actually build wrong, vulnerable applications.

Yeah. I really appreciate that because... Especially now we're we're in the heat of our call for submissions for RSAC conference twenty twenty seven, and we're seeing, you know, so much change so quickly because of exactly what you said. Like, those who don't necessarily have the technical skills are able to do, create, implement a lot of things now through the use of AI. And I'm I'm just wondering, is that why you decided to take a deep dive look into these app generator platforms? And maybe you could even walk us through your exploitation thought process.

Actually, the main reason that we targeted the app generators is because they are trendy. And the other reason is that they must produce production applications for people who have never configured a database or written a line of code. So we wanted to know what that looks like on a larger scale, essentially Internet scale, rather than a single project.

So the first thing that we did is using certificate transparency, we could basically identify nearly all published applications. And this is because each application, even on a custom domain, gets a TLS certificate. And every certificate lands in public CT logs.

We then fingerprinted each new host to confirm that it was actually built with Lovable, and then this gave us a massive attack surface that was just waiting for us. The first day was over a thousand. Now you mentioned earlier it was about 40,000, but I think today it's already over 100.

And so were you looking just at Lovable or did you also look at like Bubble, Base 44, EpiPie, like all of these other different ones or were you just looking specifically at one?

So our main campaign focused on Level because Got from our research, this is the most used platform right now. But, obviously, we looked at all the others, just not in...

On the same scale. To add on what Ari said, we actually knew and we know, everyone knows that that AI hallucinates. So imagine someone who doesn't know about security and not even about web development starts to build production ready applications. So he gives all the authority to the AI to decide about the security model, the threat model, and AI hallucinates.

So that was our pivotal moment when you...

When we understood that, okay. We are security savvy, and we can build with AI, but we can actually spot when AI is making things we don't like from security perspective. But most people who use these platforms doesn't know that. So to us, actually, using certificate transparency, penguin printing, and finding all instances of Lovable and finding these tiny loopholes that people made because AI hallucinated was really fascinating for us, and that's what actually initiated the whole process.

It is fascinating. So, Ari, continue to walk us through this this process. Sure.

So as soon as we got our attack surface, we've seen that most of them are wired to that same back end as a service super base and that the front end bundle ships the project URL and the public key, and that is by design.

So the only thing that was protecting the data is raw level security.

And we ran a read only check on whether an anonymous visitor could read tables, and that's where you can see the pattern. Sometimes, depending on basically how the AI decides to implement it, and, again, it changes because it hallucinates, Sometimes apps are going live with data that is often sensitive, open to anyone to read. And sometimes there were even cases where we could write data as well, leading to basically changing the website content or even leading to potential cross site scripting on that website.

Those apps that were exposing sensitive data, like, how often were you seeing that that was the case?

Very often. Yeah. Really?

Yeah. Pretty often.

So it was... I mean, we can we Yeah.

We we can talk about disclosure a little bit later. But when we started to do this research, we found a lot of loopholes, as Ari just mentioned, with the super base kind of misconfiguration. So this is pretty insane because imagine you're going to a website that you just built and it has a key inside of it that anyone can find and anyone can use. Now unlike the traditional way of how websites and web application were built before that because you had your front end and your back end, and you could only actually cross to the database if you were authenticated and authorized.

So because these platforms use a SaaS, which is a back end as a service The applications from the front end side, they have to own a token, okay, was prior to that a session cookie because you've been able to actually register and sign in. But right now, you go to a web app, you can find this anonymous key. That's how superways call it, anon key. And because AI hallucinates, this key can actually allow you to read stuff from the DB or write stuff into the DB that you shouldn't, which can lead to, as Ari mentioned, XSS payloads and the exfiltration and expose of sensitive data.

And, you know, like, this type of all platforms using the same back end as a service platform, knowing that that can hit thousands of people in just one go, I do wanna look at the real world stakes.

Have you guys seen these type of attack vectors being exploited in the wild? What does that look like? And then who should be the most worried about these discoveries?

So we actually cannot say that we are a 100% sure we we we didn't see exploitation in the world. But it's always worth remembering because we as security researchers, we get that question quite a lot, and we are not so different from the hackers. We are actually the good guys who mimic their steps. So the answer is no.

We didn't see any exploitation in the wild, but it's pretty obvious to to anyone in the hacking industry that this is something that you should pay attention to. Now imagine Superbase boomed really strong. Everyone started to use it. Now for for hackers, for security researchers, that's the holy grail because once you know there are a lot of tools, a lot of platforms relying on the same single one platform, you just focus all of your energy and time finding a vulnerability, just one in this single tool, in this single SaaS back end of the service named Superbase.

Once you found one in Superbase, you can actually hack all of them. So in addition to that, I would say, who should be worried about this? I would, first of all, want to talk to the security teams within small businesses, within enterprises. Because the fact is, and we are seeing it within our customer base and within our talks with clients and just general research, Enterprise personnel use these platforms.

So imagine you're in a big enterprise, and the enterprise used to ship code reliably in a government way after code review, after a bunch of different CICD pipelines that meant to enhance and protect from vulnerabilities. And right now, it's just a few words, chip, click by someone who doesn't know anything about tech in terms of software development and that doesn't know anything about security. So the people who should be worried are security teams that have zero visibility over the people within their enterprise, within their organization that use Lovable or any other platform, integrate it with sensitive data, and actually ship it to be used internally or externally by customers.

Yeah. So it sounds like there are a lot of different steps that folks can take. Right? You know, first of all, you wanna make sure that your application is more as secure as your configuration. Right? And so you wanna make sure things are configured properly, but it sounds like also there's a nonkey publicly sharing data is risky. So what's one actionable takeaway you both can provide to our listeners that can... That they can use to protect themselves or their organizations today?

Honestly, the thing that I would focus about most is, as we said earlier, the certificate logs are public, and security teams should be using them to alert and monitor new certificates for company subdomains, fingerprint them at scale. And then this way, you'll be able to catch the AI built app someone from marketing shipped last night before the attacker does.

Basically, we think that people should actually treat their environment the same way as hackers treat their environment. So, you know, the whole phenomenon of purple team and attacker mindset that we are talking about for the past twenty years probably, this still works and it's still proven. So we think that actually if hackers use this thing called certificate transparency and it allows them to build a huge bank of leads, the security teams can do the same. And it could be really helpful to understand, okay. That's the surface that hackers are seeing right now. Let's deal with that first. Because if that's what they're they're seeing, let's tackle this first.

Now to the other point, I think one of the most pivotal takeaways is these applications, these platforms were built to consumers in the beginning, and they actually used it as, quote, unquote, toys. And what I mean by that is many of these platforms, they actually ran competitions where people build crazy stuff like the seven year old who build this amazing chess game or the 14 year old who build his own ecommerce platform And so on and so forth. So it was...

People actually use it as a toy, as a very unique toy that can actually build for you everything, which is crazy. But then we move forward to right now, and it's not a toy anymore. It's actually something that is being used by Fortune 500 companies and can actually leave them exploited and exposed. So I think one of the most important thing is to treat these app generators as something that should be governed properly Like any other method within the company.

Because, traditionally, companies right now know how to ship code safely in the traditional way. You know, people push code to GitHub, so on and so forth. GitHub has CI, CD pipelines. We have code reviews.

Everything is already known, but this disrupts everything. So I would say the most important thing I can say is to actually take these platforms, create visibility for them, put them under SSO, and decide who is allowed to actually publish stuff or integrate these stuff with sensitive data. Now this is not super easy to do because every platform is different, and not all platforms actually support enterprise level usability. But there's a lot of different ways to help customers.

There's a term right now that we use called ShadowAI, which means we can actually help customers find these stuff even if the platform doesn't support them in terms of enterprise subscription.

But, again, just to summarize, this is the main thing I think organizations and security teams and even consumers, although they haven't... They're having fun with it, doesn't thinking... They're not thinking about security too much, should take from this episode.

Yeah. I think that's great advice. I think, you know, at the beginning of our conversation, we were talking about, you know, maybe it's the marketing or the HR person that's just sort of pulling together or building out this application using these tools and visibility into what's going on across the entire enterprise is really important and who is doing what. Like, that's the shadow AI you talk about Guarding that key, man, it sounds like that a nonkey really needs to be secured.

And and then auditing the AI generated code. Right? Like, making sure that there's a review of the code, I think, are all really great advice. Roy and Ari, thank you so much.

This is such an interesting conversation. So glad to have you here with us. Listeners, thank you for tuning in. Please keep the conversation going in our RSAC membership platform by visiting 1rsac.com/membership,

and be sure to check 1rsac.com for new content posted year round. Until next time.