

Cybersecurity for AI-era infrastructure—“everything hackable will be hacked.” The guest argues outages and breaches may be linked to attackers exploiting interconnected systems, including via agent-driven lateral movement and supply-chain compromises. He claims companies must triple down on defensive security, guardrail coding agents, and audit code before deployment; otherwise agents can uncover backdoors and secrets that enable cross-company attacks.
Vercel Security Review reportedly found 1,000 critical/high vulnerabilities in customer code within 24 hours; third-party front-end pixels can lead to hacks when embedded unsandboxed code is compromised; Hugging Face was hacked and used open-weight models to analyze an “agent escaped the sandbox” scenario.
Gene (Vercel CEO). Background: infrastructure/cybersecurity expert; leads Vercel’s coding-agent security approach (Vercel Security Review; open-source DeepSec).
Written by AI. May contain mistakes.
Discussion on the current state of cybersecurity and its implications for infrastructure.
“And so the cautionary tale for companies is you have to be investing and tripling down on your cyber security.”
Exploration of the importance of code auditing and the risks of third-party dependencies.
“And I can't tell you FAA planes being grounded is related to this, but there's definitely a world.”
Talk on the need to rebuild software systems and the role of cybersecurity AI.
“website and so i think um going back to like why you have to audit the code a coding agent can tell you today, hey, that's a bad choice.”
Insight into evolving hacking methodologies and the value of AI in defensive strategies.
“So do you expect more companies to build entirely new systems and then migrate over?”
Analysis of NVIDIA's ecosystem approach and its implications for global AI development.
“But also they found a new reward for hacking.”
Discussion on America's role in leading the AI revolution and its implications.
“And so overall, I think the world that NVIDIA was pushing for is overall very, very positive.”
Exploration of high-profile hacks and the necessity for improved cybersecurity.
“My hottest take perhaps is now becoming lukewarm as this conversation evolves.”
Introduction to DeepSec, an open-source cybersecurity solution.
“and we're sharing our resources and what we're doing with the world and we're investing in the open.”

Get the gist, explore an idea or ask about something said.
On Sourcery, what are the key points of this episode?
Automatic transcript. May contain errors.
0:00Gene, welcome back to Sorcery. It's great to be here. There's a lot going on in the world right now, and you seem to be an expert on all things infrastructure, and you've seen a thing or two. So, what is going on? Yeah, anytime you hear about these broad outages, the very first thing I think about is are these people taking their systems offline because they just got breached, right? Cyber security right now is, I think, the number one threat to the stability of the world's infrastructure. It used to be that if you had a vulnerability, it took a very dedicated adversary to exploit it. Now we all have tokens.
0:39We all have agents. Agents have superhuman hacking abilities. And so the cautionary tale for companies is you have to be investing and tripling down on your cyber security. You have to be putting these tokens towards defensive cybersecurity use. Anytime you hear about, you know, like, oh, company X got breached, you should also be thinking about, well, if that company gets breached, what about the data, the secrets and the documents that the agents can find in there that allows them to pivot and move laterally and hack other companies? And so what we've been hearing the last couple of weeks is all a network of hacks that sort of have been happening.
1:24some of which I've heard from first-party sources like, oh, yeah, the hackers got into that, and that allowed them to move into that, and that allowed them to move into that and crossing company boundaries. Yeah, so I advise people to think also about their personal cybersecurity. You know, anytime you have an opportunity to turn on multi-factor authentication to, you know, the way that a lot of these hacks happen is you download software or the supply chain gets compromised, and the hacker pivots or the agents pivot from a computer to another computer. And so what we're doing at Vercel is we're giving people turnkey solutions so that when they produce code, we can audit it with agents.
2:08The agents can find vulnerabilities before they create the gap that other agents out there in the wild can exploit. So we announced Vercel Security Review, I think it was last week, Within the first 24 hours, it identified 1 ,000 critical and high vulnerabilities in code that customers were attempting to deploy. And so what you have to think about is that you're using cloud code, you're using codecs. You have to be really paranoid that that code, if it's not audited correctly, might contain the future backdoors that other agents can attack. And so we all sort of need to start investing. My prediction is that this is going to be the defining topic for all of the software industry.
2:58In a way, you could argue like we've solved coding. Coding is now virtually free, abundant, but we really need to solve cybersecurity if we want the infrastructure to remain stable. And I can't tell you FAA planes being grounded is related to this, but there's definitely a world. You should think about it because the networks are so interconnected that if your infra gets breached and that can have all of these cascading effects, as we roll out AI more broadly, this is to be top of mind. Talk about how simple this can be. It could just be like one pixel on a website. Yeah. So for 10 years now, we've been advising companies to take their front-end engineering very seriously.
3:43Our first product that Vercel brought to market was called the front-end cloud. And one of the things that we helped people do was not only improve the performance, but to actually own every pixel, meaning all of the code is of your own provenance. It's like single origin coffee. We helped companies have single origin websites. But a lot of websites out there are actually made as a juxtaposition and tech debt pile on top of tech debt pile on top of tech debt. And so a lot of really high-profile hacks that have happened recently is this website had a pixel loaded from this other third party and that third party was not sandboxed and guard railed correctly and now you're going to chase.com and it's not because chase got hacked it's the third party that got hacked happened to be embedded there on the on the front end of the website and so i think um going back to like why you have to audit the code a coding agent can tell you today, hey, that's a bad choice.
4:42Just are you actually putting it as part of your software development pipeline? Are you willing to hear the bad feedback of like, hey, that's dumb. You should not do that. And so we have the solutions. We have to like as an industry broadly disseminate them. And the reality is we're dealing with massive amounts of tech debt. Our infrastructure and software systems have been built from the ground up by humans for decades. It's actually time to start rebuilding a lot of these foundations. We can rebuild them in memory-safe programming languages. We can rebuild them in ways that are almost like provably unhackable.
5:20And I think so we're going to have some pain in the next weeks and months as we hear more about these breaches. I think it's very likely that they're going to happen. But I think on the other side, as we start deploying this cybersecurity AI systems, we're going to end up in a very, very strong position. I mean, I think that's a good point. So do you expect more companies to build entirely new systems and then migrate over? Because migration doesn't take like a whole team of consultants and like multiple months anymore. Not anymore, right? So obviously I'm CEO of Resell, so I kind of see this every day.
5:56But when an enterprise buys our platform, we help them deploy new software that they can now produce infinitely faster because they have coding agents. and the default way that enterprises roll out of their cell is behind the firewall. So literally, the attackers can never even reach into these private networks. Now, conversely, the risk that exists with a lot of the agents is that they're going to throw your data out into the public internet because software, like I said, is now basically free. It's so easy to generate. What you need to do is make sure that it's properly governed so it doesn't accidentally, kind of like OpenAI accidentally hacked Hugging Face, you don't want your agents to accidentally yeet out your data into the internet.
6:42And so I think our solution is embrace the coding agents, guardrail them, deploy mostly internal software, and if you do need to deploy to the public internet, get that extra set of AIIs that are warning you, even before the code goes out, hey, like you could be improving your security posture. So do you think what's going on right now is less of a tinfoil hat moment and something that is just obvious and inevitable? I mean, if you're running critical infrastructure, your default attitude has to be like three layers of tinfoil around your head. Like, Because the gaps exist, and as you get deeper into this labyrinth, you realize there's risk everywhere.
7:32And so the question is, how do we mitigate it as fast as possible? And the answer is, we have to deploy AI. So we're all using AI in creating new apps, sites, images, whatever. We now have to deploy AI towards protecting these systems. I think what we're seeing now is the models have gotten really capable of cybersecurity. And so the bad actors have found a new methodology for hacking. But also they found a new reward for hacking. So the targets of these hacks are all companies that have access to AI tokens. And you might have heard of these distillation campaigns and distillation attacks that are happening.
8:17So if you hold a lot of AI, whether it's North Korea or what have you, they're going to try to hack you because then they can get AI, which is now almost like the new currency. Tokens are very precious. They're fungible with U.S. dollars, right? And so there's black markets not only for data and these data boundaries that exist, but also for tokens. And so the reward for hacking has now increased, essentially. What do you think is going to happen with the open-weight world? What I've seen is every company on the planet needs to use AI in order to basically think like a hacker. The best way to find vulnerabilities in your code is to audit it like a hacker would.
9:04And that's where open-weight models can actually be incredibly helpful. Because the coding agents that you use to produce the code, they miss stuff. And sometimes they even have guardrails where they don't allow you to look deeper into what could go wrong. And so I actually like a world where all of the models are almost like pointing guns at each other. Like you almost want an agent that distrusts what the other agent produced. So if you're familiar with Vercel, anytime you ship code or you push a Git commit, we build it and deploy it. And so in that pipeline, we insert an agent that uses a combination of proprietary and open models that basically takes this very sus attitude and says, yeah, I don't trust this code.
9:51Where could it go wrong? And so also when Hugging Face got hacked, they actually used open-width models in order to make sense of the attack. They were like, okay, the agent escaped the sandbox. Help me understand the surface of attack here. And so we've seen a lot of data and anecdata of companies using open-width models for defensive purposes. I think that's a really good thing. Should we talk about NVIDIA? Are you at all involved in this? What I love about NVIDIA is that they've always been a company that builds an ecosystem. They've never tried to own the full pie. In fact, when they've gotten into models, they started open sourcing models.
10:37like Nemo Tron. So NVIDIA, I think, is extremely spiritually aligned with the mission of Vercel, which is to help customers own their intelligence and to own their agents. Of course, we collaborate with all of the labs, et cetera, but I think the best outcome for the world, for the free world, is that we all have free access to intelligence. Free as in low cost and as in I can modify it. You know, Vercel was born out of open source. So many companies run thanks to the success of open source software. You know, famously, Facebook started out by getting runtimes and frameworks and programming languages which were open source.
11:21So the next great American companies are going to be born out of ecosystems. We need to maintain open ecosystems. And I think Jensen was doing a phenomenal job at building an open ecosystem. and I think a lot of people are very invested in his demise because of that. Now, obviously, we can all have the moral high ground of like F. Jensen. He was doing a lot of wrong things, clearly. But there's also, I think, a little bit of a fog of war. It's like, why are you out to get Jensen? Is it this? Is it that? And what is your motivation? I think people need to study the matrix a little bit more deeply.
12:00What should we be studying in this matrix? Well, I mean, what are the reasons that you would want NVIDIA to fall, right? Well, one is that if you're building a fully proprietary ecosystem and NVIDIA is trying to fund the open-weight alternative, I think that you have a reason to, at the very least, be like, okay, good riddance, Jensen. Sorry that happened. Bye. Into Oblivion. Obviously, there's the classic people that are dogpiling after any kind of mistake is made. I think he's made a lot of mistakes. And, like, there's always a peanut gallery of people that are like, ah, like, I'm perfect. You're not.
12:37Everything in my personal life is 100 % perfect, and you are an asshole. And so I think there's an element of that. And then there is legitimately, I mean, like, this is the most valuable company in the world. Everyone wants a piece of this. Some people are short. Some people are long. And so I think anytime that something like this happens, it's never about just one or two stories that break out about what the CEO was doing. We're talking about the most valuable company in the world. And we're also talking about geopolitical issues, right? And so it's true that NVIDIA was building an ecosystem that's open in every sense of the word.
13:16They sell internationally. You know, Jensen in interviews was very clear that he wants a world where China depends on the U.S. As opposed to a world in which China built something that could arguably be more formidable than the U.S. Maybe a cautionary tale here is look at the success that China has had with video models. They're mogging like Seedence. Even the success of TikTok, they're mogging. And so it's good, I think, and healthy to be in a position where they depend on us rather than like they have first party everything. First party data sets that could be larger than ours. If you want to create a great video model, it turns out to be it's pretty helpful if you have billions of people generating video data every day and your TikTok.
14:06Right. And so overall, I think the world that NVIDIA was pushing for is overall very, very positive. and at the same time, of course, want America to win and continue to be the spear of the AI revolution. I think it's actually existential for America to be the spear, as you've seen with cyber warfare. You want the best weapons, and you don't want our adversaries to have the best weapons. Damn. Okay, I said it, but we're going to have a very hard question now. This is the final question. What is your hottest take right now? My hottest take perhaps is now becoming lukewarm as this conversation evolves.
14:48There are going to be a lot more high-profile hacks. No. We put out an essay that I recommend everyone to read, which is this hot take in essay form called Everything Hackable Will Be Hacked. Everything out there that has a potential to be hacked can now actually get hacked. And so it's probably not a hot take, but everyone needs to really shore up on their cybersecurity. and we're sharing our resources and what we're doing with the world and we're investing in the open. We actually open source a cybersecurity harness called DeepSec. DeepSec allows you to spray raid and kill up all of the bad stuff on your infrastructure.
15:31But there's likely going to be some short-term pain. I can't tell you of what kind. But remember when I was at an airport one day and CrowdStrike had that famous big downtime and everything was a blue screen. There's probably a scenario where stuff like that happens, sadly, where something goes wrong and a lot of critical systems go down and we have to really be tinfoil, hat wearing, and fight hard and invest to mitigate those risks. Good hot take. All right. G, thank you so much. It's great to see you. Hopefully everything goes well in the world.
Guillermo Rauch, Founder & CEO of Vercel, joins Sourcery at TomorrowX to discuss the growing threat of AI-powered cyberattacks, why cybersecurity could become the defining challenge of the software industry, and his prediction that "Everything Hackable Will Be Hacked."
As coding agents like Claude Code and Codex make software development faster and cheaper, Rauch warns that the same technology is giving attackers unprecedented capabilities to exploit vulnerabilities across interconnected systems.
Rauch explains how Vercel is using AI agents to audit code before deployment, including its new Security Review product, which identified 1,000 critical and high-severity vulnerabilities in its first 24 hours. He discusses the risks of compromised software supply chains, the growing black market for AI tokens, the importance of open-weight models for defensive cybersecurity, and why enterprises may need to rebuild decades of legacy infrastructure.
We also discuss Nvidia and Jensen Huang, the battle between open and proprietary AI ecosystems, the geopolitical implications of US-China AI competition, and why Rauch believes open access to intelligence is critical to America's technological leadership. He closes with his outlook on the next wave of high-profile hacks and what companies should do now to protect their systems.
Molly on X: https://x.com/MollySOSheaGuillermo on X: https://x.com/rauchg
Brought to you by:• Brex—The modern finance platform, combining the world’s smartest corporate card with integrated expense management, banking, bill pay, & travel. https://brex.com/sourcery Follow Sourcery for the latest updates! https://www.sourcery.vc/TIMESTAMPS:00:00 Guillermo Rauch on Global Outages & Cyberattacks00:46 AI Agents & the Growing Cybersecurity Threat01:57 Vercel Security Review Finds 1,000 Vulnerabilities03:33 How a Single Website Pixel Can Enable an Attack04:53 Rebuilding Secure Software Infrastructure05:50 Enterprise AI Agents, Guardrails & Data Protection07:37 AI-Powered Cyberattacks & Tokens as Currency08:46 Open-Weight Models for Cybersecurity Defense10:15 Nvidia, Jensen Huang & Open AI Ecosystems12:00 Who Benefits From Nvidia's Decline?13:05 China, Geopolitics & US AI Leadership14:35 Everything Hackable Will Be Hacked#podcast #investing #technology #venturecapital #entrepreneur #startup #siliconvalley




