Skip to content
a16za16z

AI Is Learning to Hack. Faster Than We Expected.

Joel De La Garza is joined by Dylan Ayrey, co-founder and CEO of Truffle Security, and Feross Aboukhadijeh, founder and CEO of Socket, to discuss one of the biggest shifts happening in cybersecurity: AI models are no longer just finding vulnerabilities—they're exploiting them. As frontier models become increasingly capable of hacking, software security, supply chain attacks, and cyber defense are entering a fundamentally new era. The conversation explores AI-powered hacking, software supply chain attacks, leaked credentials, zero-day vulnerabilities, package manager security, and why the path of least resistance for increasingly autonomous AI systems may also be the most dangerous. They also discuss what enterprises, developers, and the open-source ecosystem need to do to adapt as the gap between vulnerability discovery and exploitation continues to shrink. Timestamps: 00:00 - Intro 00:49 - Models Are Escaping Their Cages 01:28 - Opus 4.6 Committed a Felony to Complete a Task 05:20 - The Apache Foundation Key & the Path of Least Tokens 09:19 - How the Labs Trained Models to Hack: Reward Functions & CTFs 11:45 - A Quarter Million Live Keys in Hugging Face Training Sets 13:02 - The npm Worm: Hundreds of Repos Breached During Black Hat 16:55 - npm's Nuclear Option: Mandatory 2FA for Every Publish 21:06 - 2026 Is the Year of the Software Supply Chain Resources: Follow Dylan Ayrey on X: https://x.com/InsecureNature Follow Feross Aboukhadijeh on X: https://x.com/Feross Follow Joel De La Garza on LinkedIn: https://www.linkedin.com/in/3448827723723234/ Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg 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. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

Joel De La GarzahostDylan AyreyguestFeross Aboukhadijehguest
Aug 7, 202623mWatch on YouTube ↗

EVERY SPOKEN WORD

  1. 0:000:49

    Intro

    1. JG

      Models are actively escaping their cages, going out on the internet, and doing pretty nasty things.

    2. DA

      Recently, we found an API key that had been leaked on the internet that had administrative access to the Apache Foundation. Interesting thing about cybersecurity in particular is the reward function is incredibly well-defined. Get access to the data. Did it get access to the data? Reward the thing.

    3. FA

      For a long time, people had talked about this concept of an npm worm, this idea that someone could backdoor a package, get developers to install that, and then you could use the access stolen from those developers as they install it to self-propagate the worm.

    4. DA

      If the labs are making it fundamentally easier to break into supply chain, do you think the labs have a moral obligation to fund some of the problems that they're causing?

    5. JG

      I think it's really strange that they're not letting blue teams get access to these tools, but ...

  2. 0:491:28

    Models Are Escaping Their Cages

    1. JG

      Awesome. Hey, thank you so much for joining us. We've got Feross and Dylan here from Truffle and Socket. Uh, it's great to have you guys on. This has been probably one of the most interesting weeks, if not the most interesting week, in cybersecurity. Uh, not because of the Black Hat conference, which is usually the cause, but because we've now seen several instances where models from not just one provider are actively escaping their cages, going out on the internet, and doing pretty nasty things. And I think, Dylan, three months ago, I remember a blog post we, uh, [laughs] we col- lightly collaborated on together. Um, and, and you had found a number of these issues with earlier models, right, that were less sophisticated.

  3. 1:285:20

    Opus 4.6 Committed a Felony to Complete a Task

    1. DA

      Yeah, we looked at Opus 4.6 and some of the other frontier models at the time. Given the models a very simple task, there was a barrier which prevented the model from accomplishing the task unless it went and committed a felony and hacked into a system to accomplish the task, but it wasn't instructed to do so. And we found more often than not, it would do the SQL injection, it would commit the felony, and it would do what it needed to do to accomplish the task.

    2. JG

      Mm-hmm.

    3. DA

      I think when it comes to alignment issues, no one needs to worry about these models making it materially easy to build nuclear weapons, because you need to procure fissile material to do that. It's not, it's not gonna make it easier to build weapons. Everyone needs to worry about these models making it materially easier to hack into things. The bar previously was just subject matter expertise.

    4. JG

      Mm-hmm.

    5. DA

      And now the models have the subject matter expertise. They were specifically trained to have the subject matter expertise, and they're just making it materially easier to hack into just about anything that you can think of using the fundamentals that we've been talking about for years, but previously it required a subject matter expert-

    6. JG

      Mm-hmm

    7. DA

      ... to, uh, risk, like, going to jail for hacking things.

    8. JG

      Mm-hmm. DEF CON was always famous for people, for attendees getting arrested at the conference, right? [laughs]

    9. DA

      That's, that's absolutely right. But that was, I mean, that was a barrier, right?

    10. JG

      Yeah.

    11. DA

      For better or worse, like, that prevented the subject matter expertise from, from hacking into things because they were worried about being prosecuted. Um, the bar has now fallen to just asking the model, which has specifically been trained to hack into things-

    12. JG

      Mm-hmm

    13. DA

      ... to hack into things. Um, so, so that's a concern. And then the other concern is when they're incredibly goal-oriented to accomplish tasks-

    14. JG

      Mm-hmm

    15. DA

      ... um, and, and one of the tools at their disposal is cybersecurity expertise, um, they will do the path of least resistance to accomplish the task.

    16. JG

      Okay.

    17. DA

      And that includes drawing on their cybersecurity expertise.

    18. JG

      Well, and it seems like... And, you know, the classic, the classic, the classic saying is that, "Don't pick the lock if the door is open," right? I think that's, uh, from the very beginning of, of, of the security world. Um, so it's always been sort of like to, to, to, to, to go in level of difficulty from easiest to most difficult. And it seemed like initially these tools had a very finite scope of, of techniques that they would use, and it seems like they've expanded. And I think with this test, Feross, it was interesting because they now have seemed to have escaped from just doing things like SQL injection to actually, like, trying to take over packages and do social engineering.

    19. FA

      Yeah, it's really interesting to see how, um, just like humans, um, the, the models are, you know, easiest path into a company. And I think that, um, that now has become the software supply chain.

    20. JG

      Mm.

    21. FA

      And so just like, you know, a human hacker would, um, they're gonna pick the easiest way in, and the lowest hanging fruit now has become, you know, just publishing malware to, um, to public registries because they know that there's no vetting happening and, you know, developers are likely to install them. Um, I thought it was pretty interesting, there was, um, research published recently about, um, uh, what they're calling kind of like u-universal, um, uh, typosquats or universal hallucinations-

    22. JG

      Mm-hmm

    23. FA

      ... where all the frontier models all, um, have the same, um, make the same mistake and sort of assume there are certain packages that exist that don't, um, despite, like, the, those models coming from different companies. And so, you know, uh, I, I think there's just, there's, there's been kind of enough, um, uh ... I guess, yeah, there's just like the l- the low hanging fruit of the supply chain has just become kind of s- so appetizing that even the models are trying to get in on, on the action. Um, and, um, I think the AI is, like, not only kind of attacking, but it's also kind of the way in a lot of times on the d- on the, on the, on the kind of developer side because, um, we, we see so many, you know, even non-developers using these tools to, to, to inadvertently write code or-

    24. JG

      Yeah

    25. FA

      ... you know, code comes in, uh, packages come in in order to kind of, um, build, uh, you know, graphs or visualizations or different, different things that, you know, folks are doing with these tools. Uh, and, uh, and, and, and no- it feels like no one really knows what's being installed and what's going on. And, you know, this is just basic stuff. This isn't like f- I mean, it sounds like it's sci-fi stuff, but it's really just basics. Like, what software are we using? How are we vetting it? You know, uh, just the basics of, of-

    26. JG

      Yeah

    27. FA

      ... computer security.

  4. 5:209:19

    The Apache Foundation Key & the Path of Least Tokens

    1. DA

      Can I touch on the supply chain a little bit?

    2. JG

      Mm-hmm.

    3. DA

      So recently we found an API key that had been leaked on the internet that had administrative access to the Apache Foundation. And it's like if you're in the shoes of the model and your goal is to get access to some data-

    4. JG

      Mm-hmm

    5. DA

      ... um, certainly backdooring Apache is a pretty effective way to do it.

    6. JG

      Mm-hmm.

    7. DA

      And, like, to get access to Apache, are you gonna use the secret that just allows you to directly log in, or are you gonna burn tokens and tokens and tokens on trying to find a zero-day? They're optimized to use the path of least tokens to accomplish their goals.

    8. JG

      Mm-hmm.

    9. DA

      Of course, they're just gonna use the secret that's laying out there [laughs] in the open to accomplish what they need to accomplish. And so, yeah, I think, uh, supply chain and secrets, um, are and have been the path of least resistance and will continue to be so as the to- the, uh, the models are incentivized, uh, to use fewer and fewer tokens to accomplish their goals.

    10. JG

      Well, one of the things ... And, and I think that's absolutely right, and I think it's, it's, it's that sort of chain of escalation, right, where if one thing fails, try another. And like at the top of that pyramid, right, the top of the hacker ecosystem is the zero-day vulnerability, right?

    11. FA

      Mm-hmm.

    12. JG

      It's basically finding a vulnerability that can be exploited in a product that everyone uses that you can use to basically unlock all the corporations.

    13. FA

      Mm-hmm.

    14. JG

      And one of the really fascinating things about the breach disclosure that was made was that there's, there's an incredibly popular CI/CD tool that I think every enterprise uses that this thing just spat out a zero-day for.

    15. FA

      Mm-hmm.

    16. JG

      Right? And like, I guess, like what, what sh- like, and that's like just such a critical point in the supply chain that everyone should be thinking about. Like kind of how are you thinking about that? Like that's, that's really difficult.

    17. FA

      Like the zero-day creation piece specifically?

    18. JG

      Yeah, yeah. But like for, for specific parts of that, like that control the supply chain.

    19. FA

      Yeah. Well, I mean, the, the whole world is built on this, you know, teetering infrastructure that everyone is using.

    20. JG

      It's like the classic picture of the-

    21. FA

      Yes

    22. JG

      ... the matchstick holding up the-

    23. FA

      I think-

    24. JG

      ... the complicated machine. Yeah.

    25. FA

      That, that image probably popped into all the listeners' minds right now.

    26. JG

      Yeah, exactly. [laughs]

    27. FA

      Right. And so everything from, you know, package manager registries, like we like to focus on, on that, um, because what we do at Socket.

    28. JG

      Mm-hmm.

    29. FA

      Um, a lot of those are r- you know, run by volunteers. Um, they're, they're under-resourced, you know, underfunded. Um, you know, there's lots of risk there, uh, right? And, uh, and that kind of, kind of cascades throughout the whole rest of the ecosystem. So if you look at the, you know, just the packages that we all depend on, a lot of those are single individuals that, you know ... Like, there's almost certainly, you know, we know there's-

    30. JG

      Yeah

  5. 9:1911:45

    How the Labs Trained Models to Hack: Reward Functions & CTFs

    1. JG

      from nowhere. These are learned behaviors, right? And, and, and I think what I think we're seeing is we're seeing a, a, a, a process that looks like it's been kind of maybe trained, um, or, or there's a reward structure that's been built on a bunch of these things. Like what's your understanding of how they're figuring this stuff out? 'Cause it, it seems like they know what they're doing, like they've been taught to do this.

    2. DA

      Yeah, I mean, if a lab tells you that this is an emergent super intelligence behavior, they're just lying to you.

    3. JG

      Yeah.

    4. DA

      And you can read their own safety reports-

    5. JG

      Mm-hmm

    6. DA

      ... to see exactly how the models are trained-

    7. JG

      Mm-hmm

    8. DA

      ... and exactly how they're testing these behaviors. I mean, the interesting thing about cybersecurity in particular is the reward function is incredibly well defined.

    9. JG

      Mm-hmm.

    10. DA

      Get access to the data. Did it get access to the data? Reward the thing.

    11. JG

      Right.

    12. DA

      And so when they realize that, like the number of problems that have that well-defined reward structure basically defines how we do reinforcement learning, and they want to find as many problem spaces that they can do reinforcement learning on. And so it was a prime candidate for them to come in and give it, uh, CTFs and give it like cybersecurity challenges where they say, "Okay, get access to this thing and do whatever hacking you need to do-

    13. JG

      Yeah

    14. DA

      ... to accomplish the goal." And one thing-

    15. JG

      And then they've essentially been buying pen testing data for the last four years, right? Like-

    16. DA

      Um, that's, that's a piece of it. The other piece of it is-

    17. JG

      And then the capture the flag contests-

    18. DA

      The other-

    19. JG

      ... and all those sorts of things

    20. DA

      ... the other thing is it's just not difficult to construct a challenge.

    21. JG

      Yeah.

    22. DA

      Just, d- you know, even if there is no known exploit, if we're talking about zero-days-

    23. JG

      Mm-hmm

    24. DA

      ... you put a piece of software between the model and, and, and some data, and you say, "Get access to the data."

    25. JG

      Yeah.

    26. DA

      And then, you know, if it get access, it's, if it gets access to the data, you reward it.

    27. JG

      Yeah.

    28. DA

      And it's that simple. But the other piece that they've layered on top, and this is where it starts to get really interesting, is they've started to reward the path of least tokens.

    29. JG

      Mm-hmm.

    30. DA

      And so the reason that's interesting is because for the first time it's actually able to quantifiably show us the path of least resistance for just general cybersecurity-

  6. 11:4513:02

    A Quarter Million Live Keys in Hugging Face Training Sets

    1. JG

      it. Like [laughs]

    2. DA

      That's exactly right. So, so, um, I mean, what was interesting is we were in the middle of partnering with Hugging Face-

    3. JG

      Mm-hmm

    4. DA

      ... to clean up all of the credentials that had been exposed through all of their training sets. Not Hugging Face's training, but people who hosted training sets-

    5. JG

      Yeah

    6. DA

      ... on Hugging Face. They use TruffleHog for a wide range of reasons. Um, and Hugging Face has been a great partner in getting credentials cleaned up. We targeted their, uh, training sets because we knew they had a lot of keys. Turned out there were about a quarter million live keys in their training sets-

    7. JG

      Wow

    8. DA

      ... many of which had direct supply chain implications. There was a foundational Linux library that one of the keys had direct push access to. It could have pushed malware to most machines on the planet.

    9. JG

      Mm-hmm.

    10. DA

      And so while we were in the middle of doing that, the CTO of Hugging Face shoots me a note and says, "Hey- This is crazy, but there's this [laughs] OpenAI thing that just happened-

    11. JG

      [laughs]

    12. DA

      ... and I want you to take a look at it." And sure enough, the first thing listed out in the incident response, um, although it's true it did utilize zero days-

    13. JG

      Yeah

    14. DA

      ... um, but the first thing listed out was stolen credentials.

    15. JG

      Yeah.

    16. DA

      Um, and th- that's, that's how they were trained. The path of least resistance. The path of least tokens.

    17. JG

      Password is a password is always the first step, right? [laughs]

    18. DA

      Exactly. That's exactly right.

    19. JG

      And you've, you've had your hair on fire, I think, [laughs] pretty substantially for the last, like, 18 months. Um, I think right now as we're recording this, there's currently an ongoing active breach of, of a big npm repo, isn't there?

  7. 13:0216:55

    The npm Worm: Hundreds of Repos Breached During Black Hat

    1. JG

      Something happening?

    2. FA

      It's more than just a repo. It's actually about, you know, a, a few hundred repos.

    3. JG

      Oh, wow. Okay.

    4. FA

      So, so it's a, it's a worm.

    5. JG

      Yeah.

    6. FA

      And this is one of the things that has been kind of an unfortunate innovation in the, in the malware landscape-

    7. JG

      Yeah

    8. FA

      ... uh, o- on, uh, you know, npm, is that, you know, for a long time, you know, people had talked about this concept of an npm worm.

    9. JG

      Mm-hmm.

    10. FA

      You know, this idea that, you know, if I could... someone could backdoor a package-

    11. JG

      Yeah

    12. FA

      ... um, and then, you know, get developers to install that, and then you could use the access stolen from those developers as they install it to self-propagate the worm. You could create-

    13. JG

      Mm

    14. FA

      ... you know, something that quickly takes over npm. And this had, this was kind of in, you know, passed around in blog posts-

    15. JG

      Yeah

    16. FA

      ... over the years, and no one actually kind of thought to do it until the hackers kind of figured it out.

    17. JG

      Until someone thought to do it. [laughs]

    18. FA

      Until someone thought to do it, uh, and actually-

    19. JG

      Probably using AI, right?

    20. FA

      Almost certainly, yes. Uh-

    21. JG

      Yeah

    22. FA

      ... and, and, and there's been, um, you know, uh, that, that malware, I think we, we have pretty good reason to believe that was vibe coded.

    23. JG

      Mm-hmm.

    24. FA

      Um, there's been o- one of the threat groups actually kind of posted their, you know, open sourced their-

    25. JG

      Yeah

    26. FA

      ... their kind of vibe coded, um, toolkit for-

    27. JG

      Yeah

    28. FA

      ... for others to use to be able to do this. You know, we've seen copycat attacks happen-

    29. JG

      Yeah

    30. FA

      ... since then.

  8. 16:5521:06

    npm's Nuclear Option: Mandatory 2FA for Every Publish

    1. DA

      system?

    2. FA

      Yeah, I mean, so there's been some changes, some positive movement in the community, in the ecosystem. So one thing that is positive, and it's, it hasn't shipped yet, but npm has announced that they are planning to, I think it's in January 2027, going to require, uh, human, you know, interactive, uh, uh, confirmation through 2FA before-

    3. JG

      Oh, nice

    4. FA

      ... any new publishes can happen. So that will likely kind of kill this whole worm concept, uh, completely. Um, but, um, it's gonna be super disruptive because everybody's hooked, hooked up this stuff to, you know, automation so that, you know, GitHub actions kicks off the publish.

    5. JG

      Mm-hmm.

    6. FA

      And so that's gonna break, like, a, like the whole, basically the whole ecosystem-

    7. JG

      Yeah

    8. FA

      ... when they do this, but I think it's the right call. Um, but you know, there's other ecosystems that are volunteer-run that don't have the backing of GitHub and Microsoft behind them, um, that are gonna, you know, probably not make those changes. And so I think we're still gonna see stuff like this. Um, but, uh, but yeah, we shouldn't have, we shouldn't have files on our, on our, you know, in our home folders that have tokens in them that are long-lived and that, that let you... You know, especially if you're a, you know, a maintainer with that kind of access. Um, you know, it reminds me of a, of a friend of mine. He was a, uh, he's a prolific npm maintainer, and, uh, one time, you know, we were, this is back at, like, 10, 15 years ago when I was doing th- this kind of stuff full time, and I saw him kind of type in his password. I didn't see the password, but I saw it was, it was far too short, let's just put it that way.

    9. JG

      [laughs]

    10. FA

      It was, he typed it in far too quickly.

    11. JG

      Yeah. [laughs]

    12. FA

      And it was, and I, and I called him out on it. I'm like, "Why is your password, like, six letters, man?" And, uh, you know, he said, "Well, you know," like he lives in Denmark, which is, like, a very high trust society.

    13. JG

      Yeah.

    14. FA

      And his, his kind of worldview about it was that, you know, "Well, I don't wanna live in fear and think about these things." And I'm like, "You're on the internet, man. Like, you got-

    15. JG

      [laughs]

    16. FA

      ... people are gonna, you know, people are gonna figure this six-letter password out pretty quickly." And, you know, um- There's a lot of things like that where, you know, uh, the, the folks that are the top maintainers in the world don't necessarily have the security training or even thinking about these things.

    17. DA

      Mm-hmm.

    18. FA

      And, you know, they don't have a security team, they don't have, you know, enterprise SLAs, right?

    19. DA

      Yeah.

    20. FA

      These are volunteers that are just putting code on GitHub. And so it's on actually the users, I think, the, to actually vet what they're using.

    21. DA

      Mm-hmm.

    22. FA

      It's kinda hard to say, like, you know, we just, we're a company, we just found this code on the internet and we just deployed it straight into prod and, and it's, you know, and it's someone else's fault. [laughs] You know, it's like, no, actually, you know, there's some, definitely some responsibility for, for the users of, of this software to really be, to be vetting the artifacts that they're bringing into their environments. And so I think, you know, it, there's a lot of pieces here and, you know, it's, I wouldn't wanna, um, put too much blame on people 'cause it's a hard problem. But, uh, um, but yeah, I think there's, like, a lot of places where we can do good.

    23. DA

      Well, let me ask a follow-up to that because you said, um, there are certain package managers that have resources that other package managers don't.

    24. FA

      Mm-hmm.

    25. DA

      I think one direct example of this, and I don't, uh, cast any blame on them whatsoever, they were actually great to work with, we found a caching issue in RubyGem that allowed us-

    26. FA

      Mm-hmm

    27. DA

      ... to steal arbitrary tokens and get access to arbitrary accounts, which we could use to backdoor arbitrary packages.

    28. FA

      Mm-hmm.

    29. DA

      We disclosed it to them, they got it fixed quick, but that's an example of an organization that's under-resourced.

    30. FA

      Yes.

  9. 21:0623:32

    2026 Is the Year of the Software Supply Chain

    1. FA

      Uh, I mean, at least for, for us at Socket, I think the biggest thing we're seeing is that, uh, 2026 is the year of the software supply chain. [laughs]

    2. JG

      [laughs] That you're dealing with an incident right now as the conference is happening.

    3. FA

      Yes.

    4. JG

      [laughs]

    5. FA

      And I noticed the attackers seem to pick RSA and Black Hat-

    6. JG

      Yeah

    7. FA

      ... as the times they wanna start these npm worms. 'Cause-

    8. JG

      The hacker, the, the security guys are out of the office, yeah. [laughs]

    9. FA

      Yeah. Absolutely. So I, I think that's the, that's the thing that's the biggest... You know, I, I think, you know, prior years I was having to educate people. We were, you know, always educating people about this problem, and having to explain to them, you know, this is not a theoretical risk, like, this can happen.

    10. JG

      Mm-hmm.

    11. FA

      And we'd sometimes get these reactions like, "Oh, yeah, but like, how, how likely is it really?"

    12. JG

      Yeah, yeah.

    13. FA

      And we're like, "No, it's actually very likely. Let me tell you how it could happen." And, and, uh, you know, there were many incidents to point to, but I think this year it's really broken through into the mainstream.

    14. JG

      Yeah.

    15. FA

      And there's, like, mainstream publications, you know, um, the business press-

    16. JG

      Yeah

    17. FA

      ... covering these attacks, right?

    18. JG

      Yeah. It's like front page on Bloomberg. [laughs]

    19. FA

      Yeah. Yeah, exactly.

    20. JG

      Whoa.

    21. FA

      So, so I think that is, um, you know, that is very, very good because you need that type of, you know, air cover for, like, security teams to actually prioritize and, and find budget for these problems.

    22. JG

      Mm-hmm.

    23. FA

      And so I think, you know, despite all these attacks being very, uh, you know, painful to deal with right now, I think in the end we're gonna come out really strong from this.

    24. JG

      Oh, yeah.

    25. FA

      Because we're actually gonna, gonna get budget and we're gonna get, um, the, you know, we're gonna do a lot of good this year-

    26. JG

      Yeah

    27. FA

      ... in terms of solving those problems.

    28. JG

      It's inoculation, for sure.

    29. FA

      Yeah.

    30. JG

      How about you?

Episode duration: 23:47

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Transcript of episode RtNrvPBkwfA

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.