How to Vibe Code Securely

**Feross Aboukhadijeh** (0:00)
It's the same way that probably a lot of people use ChatGPT for writing emails. You would never just send an email without reading it. That's how I'm using it. I think a lot of the people on our team are using it that way. Whereas I think that other use case where you just let it generate something and you run it without even looking at it, that's the true VOD coding, right? I do that sometimes, and I kind of almost don't want to say it on camera because it's not good.

**Joel de la Garza** (0:20)
It's okay.

**Feross Aboukhadijeh** (0:20)
It's usually for one-offs.

**Joel de la Garza** (0:21)
This is the same space.

**Feross Aboukhadijeh** (0:23)
It's for one-offs, though. It's stuff where I know that that code's gonna, I'm gonna use it once and I'm gonna throw it away.

**Joel de la Garza** (0:27)
Oh, totally.

**Feross Aboukhadijeh** (0:27)
So I don't really care, as long as it's not gonna lead to me being owned in some way.

**Joel de la Garza** (0:31)
I think we're seeing this sort of strange bimodal world emerge, where the most productive skilled people are more productive and more skilled. And then on the other end of that barbell is the least knowledgeable in this space when using this stuff become perhaps the most destructive. It seems like there's a lack of guardrails possibly for folks, even like myself, where I did CS in college and I've done it before and I've always been horrible.
But now I'm like, now it makes me good enough to be dangerous.

**SPEAKER_3** (1:01)
Thanks for listening to the a16z AI podcast. This is another short, sweet, and very interesting episode featuring socket founder and CEO Feross Aboukhadijeh and a16z partner Joel de la Garza about the intersection of AI and security. Specifically, it's about some of the new threats enabled by trends like vibe coding and overzealous use of tools like MCP, and also some best practices for helping do these things as securely as possible. As Feross and Joel explain, LLMs have been a boon for software developers and engineers in terms of productivity, and they might even help solve some age-old security concerns, including as a result of their heavy use inside products like socket. But getting there will take a concerted effort by developers, security experts, and others to harness the power of LLMs for good. You'll hear Feross and Joel after these disclosures. 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** (2:10)
You're one of the first, I think, security founders that I talked to that was actively engaging in using Copilot and then Cursor, and then sort of really leaning into this technology. And so maybe it'd be great to first start off with sort of a discussion or just an understanding of kind of like, what has your experience been with vibe coding? Kind of what are you using it for? What have you observed? And then we'd love to talk about maybe the other side of the equation that we don't typically talk about so much, which is, what are the issues that it's raising?

**Feross Aboukhadijeh** (2:37)
Yeah, well, like most companies that I think are doing it right, we're embracing this stuff as fast as we can and putting it into the hands of our team, and folks are using it to write large parts of our applications today. You know, like you said, we were pretty early on in terms of looking at these LLMs and how we could use them in enterprise applications like Socket. When we got our hands on GPT-4, that was the magical moment for us, was like, what can we do with this to go out and scan every open source package and have this human-like intelligence doing this task that would have been too much scale for a human to handle. And then obviously, all the way now through to like agents and agent decoding and all this stuff, it's super, super powerful. There are the skeptics, there are the curmudgeons that are like, you know, and I think this is true in every company. There's always some folks in certain use cases where it doesn't work the best today, but that's all gonna get better, I think, long-term. But, you know, certain types of writing, certain manifest files like Kubernetes things, certain low-level code, we see it struggle. But by and large, I mean, in a JavaScript or a TypeScript application, it's just like super good. I mean, UI code, it does really, really well. We're using this stuff as much as we can. But obviously, like, as a security company, like, the main concern that we have is like, okay, how do we make sure that we're actually writing secure code? There's multiple dimensions on which you care about this. There's like the code itself that's generated, and making sure that people are not taking it for granted, like what comes out of these models. You know, one of the things when you write code, it's similar like when you're in school and you're taking notes in class versus when you're just listening, you remember, you pay more attention when you're actually writing something yourself. And so one of the things we have seen is that when people use these tools, it's like easier to just kind of go with it and not scrutinize every single line of code and every conditional and every piece of the code perfectly as if you were writing it yourself. And so that's a big kind of scary thing for us.