Securing the Software Supply Chain with LLMs

**Feross Aboukhadijeh** (0:00)
What you find is attackers just kind of work their way up the inverse pyramid of difficulty, and they're going to start with where you leave the windows open. And on Supply Chain, we've built this system that's largely just a way to kind of go through checks to make sure that the things we know about aren't present, but it doesn't really look for these sophisticated attacks. And so now you've got very sophisticated actors using very complicated and convoluted processes to get control of things that are integral to the Supply Chain, to insert very early in the development life cycle and conquer as many things as possible.

**Derrick Harris** (0:34)
Hi, this is Derek, and you're listening to the A16Z AI Podcast, where we dig into all things artificial intelligence with our in-house team of experts, as well as the founders, engineers, and researchers working at the state of the art. In this episode, we're doing a deep dive into cybersecurity with socket founder Feross Aboukhadijeh and A16Z partner Joel de la Garza. Although the discussion hits on a range of topics, from buying patterns to when organizations might be required to use AI for security, we focus on how large language models can help secure software supply chains. Hopefully even from attacks like the recent XZutils breach that put a lot of organizations on edge. Feross kicks off the discussion explaining why these types of attacks are still so effective.
As a reminder, please note that the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any a16z fund. For more details, please see a16z.com/disclosures.

**Joel de la Garza** (1:50)
The whole security industry is pretty focused on, and maybe to almost to the point of obsession, with known vulnerabilities.
And if you look at the types of supply chain attacks we're seeing, let's take XZutils as the most recent example, you don't stop the next XZutils, you don't stop the next backdoor with a CVE database. And so I think that probably the biggest reason is just that our tools and our approaches don't even attempt to solve the problem. That's kind of what it comes down to, in my opinion.

**Feross Aboukhadijeh** (2:18)
Yeah, I think the classic saying for these sorts of things is that you don't pick the lock if the window's left open.
And so what you find is attackers just kind of work their way up the pyramid of, or I guess the inverse pyramid of difficulty, and they're gonna start with where you leave the windows open. And on supply chain, we've built this system that's largely just a way to kind of go through checks to make sure that the things we know about aren't present, but it doesn't really look for these sophisticated attacks. And so now you've got very sophisticated actors using very complicated and convoluted processes to get control of things that are integral to the supply chain, to insert very early in the development life cycle, and conquer as many things as possible.

**Derrick Harris** (2:57)
Now, so where are the locked doors? I mean, have we locked up the more obvious doors at this point, like maybe the more traditional attack vectors in supply chain is just, like you said, where the door is open or where the window is open, whatever the analogy is?

**Feross Aboukhadijeh** (3:11)
If you look at the traditional, I mean, the traditional solutions in the supply chain and broader third-party risk management market are relatively basic.
I've run third-party risk assessment teams in the past and really large ones, and you're always left with sort of like, the only way you can really validate these things is if you hire a full-time group of people to go validate this one company and their product, and by the way, you have to do this for 15,000 suppliers, right? And so, like we're still very much in the early days of this and a lot of the solutions just aren't really cutting the muster.

**Derrick Harris** (3:43)
You reference some of this, but we've had vulnerability scanning tools for seemingly forever. Like DevSecOps has been a topic for at least several years now. We've had potential problems to some of this for a long time. I guess my question is, where's the gap between what's promised with all of these previous attempts and solutions and then what's actually possible?

**Joel de la Garza** (4:01)
You know, it's easy to look for your keys where the light is.
And so that's kind of what the industry's been doing with this known vulnerability approach. It's relatively easy to determine if you're using open source code that has known problems. There are databases of these vulnerabilities that are collected and cataloged in the National Vulnerability Database or NVD, for instance. And most of the vendors out there that try to help you determine whether your supply chain is secure are really just comparing the open source you're using to this database of packages.