LLMs, Vibe Coding, and Security | Idan Plotnik, CEO of Apiiro

Idan Plotnik is the CEO of Apiiro, an application security platform built for the AI era. They've raised $135M in funding from investors like Greylock, Kleiner Perkins, and General Catalyst.

Idan's favorite books: Zero to IPO (Author: Frederic Kerrest)

(00:01) Introduction 
(00:07) How LLMs Generate Code
(02:11) Rise of Vibe Coding: Opportunities and Risks
(05:24) Debugging and Security in Vibe Coding
(09:13) Vulnerabilities Introduced by AI Code Assistants
(12:20) Security Basics for Builders Using AI and Cloud Platforms
(15:44) Security by Design and Organizational Standards
(18:08) Making Security Dead Simple: The Appiro Approach
(22:28) Winning Developer Trust Through UX and Integration
(26:59) Biggest Technical and GTM Challenges in Building Appiro
(33:55) Rapid Fire Round

--------
Where to find Idan Plotnik: 

LinkedIn: https://www.linkedin.com/in/idanplotnik/

--------
Where to find Prateek Joshi: 

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-infinite
X: https://x.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.528)
Idan, thank you so much for joining me today.

Idan Plotnik (Apiiro) (00:04.386)
Thank you very much for having me. Prateek.

Prateek Joshi (00:07.864)
Let's start with the basics. LLMs are being used everywhere in coding. So for a beginner, if you were to explain how do LLMs generate code.

Idan Plotnik (Apiiro) (00:21.848)
So the basics is that you train the model on multi-million lines of code examples. So LLM basically learns the patterns. It's not that LLM actually understand how to code like a human. And when you understand the patterns and you have a lot of features for every pattern, then

It just predicts what's next. Okay. And how the other things will look like and it evolves. And as we saw at the beginning of DLLMs, know, so most of them were not so successful when you tried, like you tried to build the code and it didn't work. So it evolved. And as you train the model on more and more.

code examples, it will just predict what next. So this is in very high level, but even today where the latest reports actually proved that you can do two to three acts more on the velocity of a senior developer. In the case of junior developer, it's not the same thing because you

need to know how to prompt or how to engineer the prompt correctly. And this is why it also depends on your knowledge as a developer and with the combination on what the model was trained on.

Prateek Joshi (02:11.47)
Wipe coding is becoming very popular and you just type in English and it's writing code, people are building apps and with that, the issue is that many people who are not engineers are accumulating what they call tech debt. So maybe before that, what is the state of play in wipe coding? Meaning what can these tools do well and where are the biggest gaps?

Idan Plotnik (Apiiro) (02:40.046)
So first and foremost, think it's just the beginning. Like we are, it's coming fast and it's amazing and you cannot stop it. Literally my kids are building apps now. Okay. And my kid is only six years old. But I think that the main value that we see on a day by day basis is for

product managers to build a prototype without the resource of a software engineer. So you can show to a customer a working product type or a potential customer, a working prototype in, you know, in hours, which things before that took you days, weeks or months. And I think

we as a company even see it as a force multiplier. So the TLDR, I think that there is a huge value in vibe coding, but not in production at the moment. So I'm talking to a lot of CTOs these days. They're not like only if it's a POC and you need to prove something.

It will get into the code base, but if not, it will stay out and it will never serve the customers in, in, of course, I'm not talking like an enterprise scale. is no way. Okay. Because it needs to go to a lot of security reviews and a lot of architecture reviews. And nowadays, when you go and you vibe coding, you don't see the source code in some cases. Okay. It depends on which.

And we see more and more vulnerabilities in these types of architecture because the team behind the company that actually provides you the vibe coding platform, they didn't think about security. Okay. They moved faster, the faster, faster. And so they see, we see databases that shares the same data of multiple customers. Okay.

Idan Plotnik (Apiiro) (05:04.642)
And we see a lot of security risks behind that. Having said that, this is a pure force multiplier for POCs, demonstrations, but not in production level software.

Prateek Joshi (05:24.328)
I, that's a great point. And I love wipe coding and I constantly wonder that, Hey, I, I'm in a lifelong engineer. when I wipe code something in, you know, say Python, it feels amazing and I can ship very, very fast. But I also tried wipe coding in languages that I've never done before. And debugging was maddening. Like, so I wonder how a person who's, who's looking at something like a new language for the first time.

Idan Plotnik (Apiiro) (05:46.734)
amazing.

Prateek Joshi (05:54.308)
Debugging could be maddening. how do you think about people who may not have used programming languages before? Now, they wipe code something, the app comes into existence, and they just can't understand it because you never had to deal with it. So debugging issues, so wipe debugging, how do you do it? And also, what specific security risks are these tools introducing today?

Idan Plotnik (Apiiro) (06:19.116)
Yeah, this is a great question. By the way, I must say that I personally, which again, I hope that my team doesn't hear the podcast. Okay. But I'm not allowed to code. Okay. I'm not allowed to code, but I am doing it. And I explicitly trying different languages that I've never developed in. And I think.

Prateek Joshi (06:31.129)
haha

Idan Plotnik (Apiiro) (06:47.704)
the value of asking an LLM model, what does this code do? Give me examples, explain to me the logic, explain to me the structure, the syntax. It's amazing. Like literally, this is a unique differential. This is, think, something that I have never experienced personally. But I think that the ability to explain.

in a very fast way and an accurate way that you don't need to know a specific language for all the bits and bytes because you will get examples of what worked, what didn't work and depends on what tool you're using, it will also go and try to understand in public data sources what didn't work to other people. So you can learn and iterate.

and ask the model to explain you things. And I think this will break the barriers of, hey, I'm a top notch Java developer. Great. As long as you understand the basics of programming and the concepts, I think to move from one language to another will be no brainer, like very easy. And your second question, I think this is super interesting because

We today in Appiro, we are analyzing millions of code repos, but not the public code repos for a lot of Fortune 500 companies. And we see we are now we are working on a research internally that will be public hopefully in the next few weeks. And because our deep code analysis technology understand

the assets behind the code. So what is an API? What is a data model? What is an AI model or GenAI framework or exit point or an entry point and on and on and on and correlate that on a graph. And then we augment more and more context. So for example, we can take a code base and say, hey, at this timeframe, someone enabled AI code assistant in this code report. So we can go.

Idan Plotnik (Apiiro) (09:13.608)
and compare before and after the behavior of the same developer. And because all of your security tools are connected to Appiro, we know if you had a growth in the number of vulnerabilities, is it repeatable? Like the same vulnerabilities by predict you did the same mistake or the model or the AI code assistant.

actually generated new types of vulnerabilities. So public research that like, know, papers that we've read saying that 26 % I think of the code is increased in a large repo, but more than 50 % of the code is actually vulnerable.

And when they said vulnerable, it's not only like the OWASP top 10 vulnerabilities, it's a new type of vulnerabilities, or I would say risks, because exposing sensitive data in an API, it's not a vulnerability, but it's a huge risk. And not putting a specific control in place, it doesn't mean it's vulnerability, but it's not the best practice.

or the security standards and policies of your organization. So the AI codices doesn't understand your security policies, your security controls outside the code. And this creates a lot of new risks. think about on top of that, I think it's publicly announced, it was announced, I saw it on LinkedIn, that a Citibank enabled 40,000

seats of AI code assist across 40,000 developers. It means that without hiring more developers, now they have between 80,000 to 120,000 developers that are writing more code and there is no way for the CISO to actually govern these types of risk. it's, again, I'm just repeating myself, it's new patterns, new APIs.

Idan Plotnik (Apiiro) (11:38.036)
new unvetted libraries and frameworks and open source dependencies. And I think this will create a tsunami of different types of vulnerabilities and other risks because the AI, if you go back to your, if we go back to your original question, the AI model only sees source code. It doesn't live outside the source code like

What's the security controls in your API gateway? Or do you have a web application firewall or other stuff? And then it doesn't understand the risk, the overall risk.

Prateek Joshi (12:20.238)
Right. And more recently in the news, the famous app T, they got somebody hacked it and released all the user IDs and the licenses and location. And people are just trying to understand what happened. And somebody posted as a joke that all of this information was available in a publicly accessible URL. So they're like,

Is it even technically hacking if it's just like going to a URL and downloading? it's like, so the point is if somebody is building a useful app with good intentions and they don't know anything about security and they don't want to deal with, I don't want to spend, you $5 million hiring a security team. What can they do to make sure that the basics are right? And also the, if I'm building something on this AI infra cloud infra platform, why aren't they taking care of my.

security.

Idan Plotnik (Apiiro) (13:19.982)
Wow. First and foremost, I will start by saying that yes, we're going into an era where I don't have all the answers. Like this is a big problem, what you are saying, what happened. And we will see more and more of that. And we, I don't want to be pessimistic for a second. I'd say we lost control. Okay. But it's a big...

attack surface that we don't have the right tools at the moment to deal with. Having said that, having said that, I do think that if you develop software, you are responsible for the data, for the security, for the compliance of what you're doing. And if not, you should not do that or you should, or users will

should not use your software. This is maybe stating the obvious, okay? But I do think that the AI code assistance and the vibe coding tools out there will need to embed security, compliance, and risk management into these tools. I can tell you

that we are working with some of the leading AI code assistants to actually embed the context that Appiro provides, which is a very unique context that is not only from the code. I can talk about this for hours, but when you want to write more secure and compliant code,

You need, you must, must, must enrich the model with more data that the AI code assistant at the moment cannot get. And some people will tell you, it's not an issue. You know what? I have an MCP server. So you can connect my cursor to 50 MCP servers and my cursor will be smarter. No, no, it doesn't work like that.

Idan Plotnik (Apiiro) (15:44.27)
You need a layer that will take the input from MCP server one, MCP server two, MCP server three and make sense of it because then you have a cacophony of context that no one can connect the dots. So it will not impact the security of the code that cursor or anyone else actually generates.

And I will add one more thing on top of that, that if we talk about Fortune 500 companies, then each one of these large companies has their own interpretation of what is secure coding. Our standards, it's different from your standards. So the model, the AI coding assistant or the vibe coding,

needs to actually understand the custom security policies and standard to be able to develop secure and compliant code based on my organization policies. If not, I can get sued. And I'm sure you know, if I'm getting sued today as a large bank and for 5 % of my revenue, this can be a huge problem for the business.

Prateek Joshi (17:10.416)
That's a great point. And let's talk about how it functions in practice. many people, they're building apps or writing code. And let's say they want to be responsible. And they're like, hey, if you give me a button, I will click it and it will audit my code and say, hey, publicly accessible URL, you're storing all this data. No way. Whoa, just stop. Let's redo it. Ideally, I want a button like that. But in practice,

Clearly it doesn't exist because people are still getting hacked, doing super obvious mistakes. So if you think about going forward to make security dead simple to adopt, what needs to happen from security companies to make sure that, hey, I know you don't know about security, you don't know the code, it's fine, here's a button and you'll be fine. So how can we make it that dead simple that people are like, okay, this is easy enough, I'll do it.

Idan Plotnik (Apiiro) (18:08.418)
So first, this is why we exist, why Appiro exists. This is our mission. And by the way, when you say at the code, you're missing one step before, which is super important, the design phase. If the design is not accurate, so the prompt will have risks before the code was written. And then you have the code that implements the design, and then you have, you know, another problem.

So what we are saying and doing and working with our customers is to do the following. We are saying when you get apparel and you understand the software architecture, you enrich it with vulnerability scanning data, and you enrich it with your compliance and security standards and policies.

And then you connect to your AI code assist. Then we enrich the prompts behind the scenes and make sure that what you're doing is secured by this from the design phase and then the coding. Now let's say that the AI coding assistant generate a vulnerability. We will also scan the output of that.

And before you open a pull request, we will tell you, okay, these are the risks. This is what you need to do to fix it. Now we are, I don't know when you're going to publish the, this podcast away. Maybe I will give you, I don't know a scoop, but we're launching a very, very interesting technology called AutoFix agent.

that takes all the data and automatically fix not in a generic way, which means if you're in an organization one and I'm in an organization two and we're both developing in Python.

Idan Plotnik (Apiiro) (20:21.518)
same framework and we have the same vulnerability, the fix will be different based on your organizational policy and your organizational security controls. And I think this is the key. Now, what I'm describing here is only for large enterprises. The large enterprises has the money, has the team, the resources to go and implement

a platform like Appiro connected to seamlessly to the AI agent, enrich it and make sure that 40,000 developers are writing secure code and govern them. Now, if I take your question to the entire world, okay, that wants to use, I don't know, a vibe coding platform that writes, you write the text.

to build the prototype, this is a little bit more challenging. Why? Because these companies do not expose the source code. There is a source code behind the scenes, but they do not give it to you. So here is a challenge. Now, if more and more people across the world will say, I'm not going to use your service if it's not secure. And then we partner.

with these types of vibe coding platform. And behind the scenes, we do the same thing that we are doing with Fortune 500 companies with these vibe coding platform. Great, amazing. Or you know how it goes. If there will be a regulation that you cannot have a vibe coding platform without having compliance and security and all these things, then...

They will must do something, okay, and integrate with companies like us or others in the market.

Prateek Joshi (22:28.4)
Right. I think this idea of security by design, meaning introduce all of this, even when doing the design phase, as early as possible so that it's easier to cache, easier to fix. And obviously, you also check the output, but sooner the better. That's great. Now, I want to shift the convo towards your customers. Basically, developers are a skeptical bunch in general. We are skeptical of any new tools.

So looking back from early days to now, what product choices or what UX choices have proven to be most effective for you in winning the hearts and minds of developers and your customers?

Idan Plotnik (Apiiro) (23:13.966)
So just to, before I will answer, just to set the expectations. It's in our case, it's not only the developers. It's a combination and partnership between application security engineers and developers. Okay. I will tell you to your question zero. Developers hates when we introduce them and new UI and new experience. So what we did...

we actually embedded the brain or the intelligence of Appiro into their tool chains. So if like I'm talking before the agent, we had a bot in GitHub, Bitbucket, GitLab, Azure DevOps, wherever you are using, And when you open a PR, you have the Appiro bot that helps you and say, hey, you're violating the

company policy, or you introduce these types of vulnerabilities and it's above critical to the business. So you cannot move on and merge the pull request into the main branch and things like that. Okay. This was the, you know what I want to call it the old fashioned way of interacting with developers. And there, based on your question on the UX experience, I think you need to be

very, very clear and we learned from all the other mistakes of other tools that interact with developers. They bombarded them with too many words, context, developers doesn't want. Give me what's the TLDR and explain me in a simple way what should I do to fix it? So we actually normalized because you can connect to Appiro.

every SAS tool, every SCA tool, every secret security tool, every security scanner that you can think of. Now, every security scanner has their own lingo, their own, I don't know, CVSS score, EPSS, descriptions. So we built a normalization layer, but the normalization layer is different to an AppSec engineer and different to a developer.

Idan Plotnik (Apiiro) (25:41.622)
It's the same normalization, but if you interact and the output is in your pull request, then we will be very, very simple, very short with an actionable thing. And we also want you to provide us, you know, a feedback if you liked it or not. And it's a constant improving, like, you know, a process where we improve based on the feedback.

directly from the developers. So one is we never took them out from their own tools. minimum data that will explain two things. What's the impact on your business? Because if you made a mistake and it will cost the business $50 million, you will listen, you will read, and you will care about your one mistake that will cost the business $50 million.

And then how to fix as fast as I can. If I don't know how to fix, who is the right person that I need to talk to? And these are the simple things that we did. And to Opsic engineer, it's a different conversation, much more complex, much more data. And let's not go there now.

Prateek Joshi (26:59.44)
Looking back, we've been running the company 2019 to now, what is the biggest challenge you've had to overcome to build the company? Let's actually, let's do part A and part B. Part A is biggest technical challenge to make it work, and part B is biggest GTM challenge, getting customers.

Idan Plotnik (Apiiro) (27:13.294)
Wow.

Idan Plotnik (Apiiro) (27:24.29)
I will start before that. No one understood what we wanted to do. We talked with customers and everything is around customers. What's your problem?

My problem is I need to deliver secure software faster. They didn't know what it actually means. And now I will answer you because this is the foundation. If your customers don't know what is the most important pain for them, like not in high level in the details. So you cannot have a deep knowledgeable discussion.

and translate that into a feature that will solve a pain for these customers. So it was a bigger problem when we started. There was no one to copy from, no one to learn from. We needed to invest a lot of time and hours with our like trial and error with the customers. Now on the technical side, first and foremost, I will say if there are appsec

entrepreneurs out there that are listening to this podcast, run away. Run away. It's the most complex domain I've ever played in and I am 24 years in cybersecurity. Okay. And this goes back to the question on the technical side. The technical side was we needed to scan code in a very unique way that no one did before because

Prateek Joshi (28:51.76)
You

Idan Plotnik (Apiiro) (29:13.39)
Everyone build ASTs. Yes, it's hard to build ASTs. It's time consuming. It's resource constraints, yada, yada, yada. But then on top of the ASTs, the abstract syntax trees, we build our own algorithms. In 2020, who thought about building your own algorithms and train it on code? We were, think, the first one.

that did it and not to find vulnerabilities to actually understand your software architecture. How many APIs do I have in my code? Who did this change? Is it a material change or someone changed the color of my login page and I don't care. Okay. So this was a very, very difficult problem to solve. We came with years, my previous company, I sold to Microsoft in 2015 and we

We were pioneers in machine learning of taking these algorithms and analyze network traffic. So we came with a lot of data and a lot of experience on how to approach things, but code is a totally different beast. And then I'm just two more points on the technical side. And then you get into a bank with 500,000 code repos. Good luck. Okay. Good luck.

Okay. And everything collapsed and, and the product was like literally broken in the first one, two, three, four, five deployments. Okay. Because we never thought about this scale. Okay. We tried with small customers. And last but not least is how you adopt very fast to customer requirements.

How do you build a platform that is adaptable to the customer requirements? I think these are the three main challenges. I can list a thousand, we don't have time, but on the business side, it's just a reflection of what I told you. No budget lines. You need to educate the buyer, okay? And say, hey, you have this problem, you're...

Idan Plotnik (Apiiro) (31:37.794)
You have the manual risk assessment questionnaires to release software to production to the cloud. Yes, it's a 10 years old process. Now you want to go and change it with your technology. You need to educate them. You need to educate them that there is a new way to prioritize and remediate vulnerabilities. So TLDR, no budget line. You need to educate the customer.

And this is why we build partnerships and we grew through partnerships. And now we have around or more than 15 % of Fortune 500 companies that are using Appiro. And they trust our deep code analysis technology, a code to runtime matching of identifying, this code asset is actually deployed here and running here and a risk.

graph that actually calculate the risk across so many data sources for every code commit in real time. And it's not only your code and not only, I don't know, your SAS finding. What's the business impact? What's your organizational policy? Where it's deployed? What's the security control? What's running? Where? And all this in near real time. And by the way, yesterday,

ServiceNow announced from their own 1.3 million followers on LinkedIn that Appiro is the industry first platform that actually updates their CMDB with code to runtime assets to allow our Fortune 500 companies to stop using these questionnaires and bombard developers or stop them from moving faster. You cannot do that. You cannot stop developers anymore.

Now they have an amazing tool called Code Assist. They're generating 3x more, which is amazing business growth. But then they have all these manual guardrails that preventing them from shipping code to the cloud. And this is what we're doing.

Prateek Joshi (33:55.768)
Right. Amazing. I have two quick final questions and do like rapid fire round 15 seconds each. All right. First question. What's your favorite book?

Idan Plotnik (Apiiro) (34:09.354)
there's zero to IPO, I would say.

Prateek Joshi (34:15.32)
Next question, which historical figure do you admire the most and why?

Idan Plotnik (Apiiro) (34:20.814)
historical figure. Wow, I have many. It depends on the context. But I will say Einstein. Like, I think, you know, he influenced every human being and I am teaching my kids the history and I'm proud of it and I learned and this

makes me happy that the human brain can invent amazing things.

Prateek Joshi (34:50.32)
Yeah.

Prateek Joshi (34:55.632)
Next question, what have you changed your mind on recently?

Idan Plotnik (Apiiro) (34:59.758)
What sorry, what I changed my mind? Oh, wow. As a philosophy, I am, I'm proactively changing my mind and MOs every time to be better and better and better and learn from others, learn from my mistakes. I think one thing is,

to talk less, listen more.

Prateek Joshi (35:34.709)
And our final question, what's your number one advice to founders who are starting out today?

Idan Plotnik (Apiiro) (35:43.371)
Number one.

Idan Plotnik (Apiiro) (35:47.402)
understand I have so many choose the right team that will fight with you when the shit hits the fan and I'm telling you the shit will hit the fan multiple times a day take take take a breath okay

Face reality, face these challenges, but without the right team, you are doomed. Doomed.

Prateek Joshi (36:24.944)
That's amazing. loved your insights. It's just the nitty gritty of what it takes to build and ship a product and build a company. So thank you so much for coming onto the show and sharing your insights.

Idan Plotnik (Apiiro) (36:37.858)
Thank you so much. last sentence that I will say is we won the RSA Innovation Sandbox, I think four plus years ago. And I said the same thing now, AI code assistant just amplifying the problem in security of software. And thank you for having me and help me like explain to the audience what's coming.

And I think we're just at the beginning of a very interesting era around security of software.