Reconstructing how OpenAI agents attacked Hugging Face

0:02Welcome to the Practical AI Podcast, where we break down the real-world applications of artificial intelligence and how it's shaping the way we live, work, and create. Our goal is to help make AI technology practical, productive, and accessible to everyone. Whether you're a developer, business leader, or just curious about the tech behind the buzz, you're in the right place. Be sure to connect with us on LinkedIn, X, or Blue Sky to stay up to date with episode drops, behind-the-scenes content, and AI insights. You can learn more at practicalai.fm. Now, on to the show.

0:41Welcome to another fully connected episode of the Practical AI podcast. This is where Chris and I don't have a guest, but we get to dive into one of the topics that's been floating around in the AI news. Maybe spend some time learning ourselves and also hopefully helping you learn and level up your machine learning and AI game. I'm Daniel Whitenack. I'm CEO at Prediction Guard. And I'm joined as always by my co-host, Chris Benson, who is a principal AI and autonomy research engineer. How are you doing, Chris? Doing good. You know, there's always so much good stuff to talk about there, but boy, do we get a good one today.

1:23Yeah, this is a multifaceted topic that, yeah, just has so much packed into it. Originally, when I saw this and what we're talking about here, we're in, for those that are maybe listening later, we're in July, the kind of end of July of 2026. and what has just happened in now time is an exploit or a hack of Hugging Face, which for those that aren't familiar, Hugging Face is kind of the online repository hub for models, data sets, benchmarks, a lot of different things for the AI community and kind of like what GitHub is for code, hugging faces for models and data sets and other things. And they reported, originally, you sent me the link, Chris, and this is when we didn't kind of know much, just that hugging face had been compromised in some way.

2:27And boy, it has developed in interesting ways as we've learned more. Just, yeah, I'm kind of amazed. I think both of us before hopping on, we were just like, Wow. So much here. Yeah. It's kind of revisiting and so much has happened. And it kind of reminds me of kind of like watching a murder mystery, you know, where it has twists and turns along the way. Something for everybody. Something for everybody there. So it's quite an interesting story. Yeah. You want to dive into getting it going there? Yeah. And just as a teaser, as we get into things, this has an element of like closed versus open models.

3:07It has an element of U.S. versus Chinese models. It has an element of agentic AI, elements of cybersecurity, all sorts of things, which is just amazing. We got something for everybody today. That's right. Yeah, exactly. Something for everybody. So just as a kind of overview, I guess an overview, and then we can dig into each one of these things. So if you haven't been following this, what has apparently happened is that OpenAI was running some of their experimental models against a benchmark, a cybersecurity benchmark, and they were powering agents as they worked on the cybersecurity benchmark.

3:54those agents escaped the test environment in which they were operating, obtained internet access, compromised Hugging Face's infrastructure, internal infrastructure, and attempted to retrieve all sorts of private information and benchmark answers from within the Hugging Face infrastructure. Which was the assigned task, we should point out. Which was the assigned task of the, yeah, so success there, I guess. Yay.

4:26but then kind of the other element of this is Hugging Face then wanted to obviously figure out what was going on. They are an AI company. They tried to use an AI model, specifically OpenAI's model, I think, to figure out what was going on by processing the log lines. And then they were blocked by the closed model provider because they were processing log lines that had malicious things in them. So then they had to spin up their own instance of an open Chinese model to actually process the logs, which all came back from an exploitation that originated from open AI. It's just like the web of things here is so interesting to me.

5:10Yeah. And it's, you know, I think one of the initial things, and things continue to evolve with the story as we want, but I think one of the initial things, like in the link that I sent you right after it happened, was the fact that, you know, Hugging Face had to go to a Chinese model that was open weight to be able to accomplish, you know, a real world task that it was trying to identify what's happened here on the cyber attack. because the Western models from OpenAI were not allowing it due to the guardrails. At that moment, I don't believe, correct me if I'm wrong, that we actually knew that the attack originated from OpenAI.

5:53I don't think so. I did not. Yeah, I did not get that in the initial. So I don't think that was really known at that point. Which is really ironic when you consider the fact that Hugging Face was initially trying to use an open AI model to discern what happened with, unbeknownst to them, was an open AI model attack. Yeah. And then had to go to the Chinese for help on this. So, yeah. I mean, I've sort of just in some ways speechless by the way that this played out. It's got all the big names. You know, I know we're nerds, but this really is quite a story. You know, if you're an AI wonk, which I'm imagining a lot of the folks listening here or watching us here are, yeah, this is quite the murder mystery for someone in this business.

6:48So keep going. Yeah, no, we can start kind of at the beginning, I guess, where it started. Always a good place to start. And I should say, it's very possible that I will say some things in this episode that are not completely, you know, people are still trying to figure out everything that happened here. I don't report to know the full details. What we're trying to do is give the picture to the best of our ability. So, you know, just for whatever it's worth, disclaimer there. But OpenAI apparently was testing some models, including, I guess, GPT 5.6. Yeah, Sol. And these are models that are more capable sort of pre-release models.

7:38So they're doing testing. And they were using a system called exploit gem, like weightlifting gem, like GYM. And to test how these models would perform when powering agents that tried to exploit vulnerabilities in code. So in an exploit gem task, apparently what an agent receives, like the input, is vulnerable source code, an input that's known to maybe trigger the vulnerability, a containerized target, and then a hidden flag somewhere in the system that it must retrieve. So it's almost like a capture the flag type of scenario. And the intended task is to convert the known vulnerability into a working unauthorized code execution, for example.

8:36And because this is, you know, on the positive side of this, a lot of people are using in cybersecurity, using AI to protect their own infrastructure, right? So just, I guess, to set that context, like this isn't a random thing that OpenAI was doing. This is very much the real world where vulnerabilities exploits are coming at lightning speed into organizations. And the only way that they can protect themselves is to use AI to help find vulnerabilities, solve vulnerabilities, etc. Which means it is, I would say, valuable. like it's uh i guess what i was trying to express is open ai wasn't trying to create the worst diabolical exploiter yeah on the on the planet uh on the on the planet let's say or at least i don't think that was their intent so so in a post mythos world mythos being the anthropic model that is not available to the public but is available to a set of organizations around the world that have particularly keen interests in cybersecurity, some of which are known, some of which are not, that I think it's well recognized now in cybersecurity that doing this process is now becoming central to the fact that you can't really defend effectively without using AI models now to do that.

10:05So OpenAI is doing something that not only are they doing, but a number of organizations are engaging in going forward. So, you know, this is kind of the new normal for cybersecurity. Yep. And because, and again, just to set more context here, this isn't just the model. There's an agent harness around this model. So going back to some of our episodes, like with Jeffrey from Noose Research, you have the model, which is kind of the brain of the agent. Then you have this harness, which is the body of the agent, which allows it to interact with an environment because a lot of the things we're about to talk about, like interactions with clusters or moving through the network or these sorts of things aren't things a model can actually do.

10:52It requires an agent harness around that model to be able to reach out to the environment and have system connections, which is why the first piece of this is so interesting, because at least partially OpenAI was trying to do a good thing in the sense that it realized, hey, we're dealing with malicious or, you know, exploits and vulnerabilities. Let's not run this agent and this benchmark in a network connected environment, meaning let's put it in some sort of sandbox. I don't know all the details of the sandbox. There's all sorts of sandbox environments. I don't know if it was just that they shut off the networking to this environment.

11:30I don't know if it was running something a little bit more strict like a G-Visor, Firecracker type of environment. I don't know all the details of that environment. But it was some sort of sandboxed environment, which in theory, it could not have interacted network wise with anything outside of that sandbox, except it could install software in the sandbox. And that's very, very significant because apparently what happened next is that the agent didn't have general internet access, right? But it could install software packages. Again, we don't know all the details, but probably through things like PyPy or NPM or, you know, those sorts of things, some sort of proxy.

12:18And what it did find was a vulnerability in that proxy environment that allowed it to execute code, which would give it arbitrary outbound network access to the public Internet. So once the model, or better put, the agent, once the agent could execute code using that proxy, which it exploited, it then could make arbitrary outbound HTTP requests. It could access services reachable from the proxy's network. So this is kind of working through that software installation or proxy, package proxy. It could read configuration, credentials, and allow the proxy to relay otherwise blocked network destinations into the sandbox.

13:12So this is so interesting, Chris. The sandboxing isolation thing, what are your thoughts? Well, I mean, there's a lot. As you pointed out, we don't know anything about it. At least you and I have not seen anything about that. there's a lot of different versions of what sandboxing can be. And some of those versions are trying to entrap, which is obviously what happened here, because there was network connection, but they thought that they had created an artificial environment that would encase that. I think you can, however, I mean, you can literally put in a sandbox that has absolutely no outside access.

13:52And that's what classified systems do and such. And so I think one of the takeaways just from this point so far that you've said is the fact that regardless of how we're thinking about models at any given point in time and people start talking about AGI and all these other things out there, none of that really matters. The point is, the agents are capable of outthinking you in this very specific task, as they clearly did based on the people running the laboratory, and find a way to do it. And that means that we are through the looking glass in a cybersecurity sense. You know, this genie is not going back in the box.

14:32So anyway, continue with the story because it's just getting good. Yeah, and maybe just on a practical note for our audience in terms of this element of it. We're all the time talking about things like a open claw Hermes agent, whether you're running like Claude code locally, whether you're building your own agents with something like LangGraph or something like this. if you're giving access for those agents through to some sort of code execution environment or where it can make system calls or network calls, you should be thinking about the sandboxing element of your agents. So this isn't just like a research sort of thing.

15:17It's very much like I have Hermes agent running on my laptop, right? And all the time here at our, like, This has been a big topic for us internally as a company. How are we doing this sandboxing? How are we protecting ourselves, but also creating a product that allows other people to do this? And so, yeah, I think if you're a practitioner and you're using one of these agents that necessarily includes this element of code execution or interactions with the environment, you should very much think about the sandboxing element. To your point, there's a whole variety of ways that you can do this from different open source projects to proprietary ones to commercial projects.

16:01But as you proliferate agents, you can't just assume that they're going to behave appropriately. They're going to try to accomplish the task, which you mentioned, Chris, the agent was trying to accomplish its task explicitly. It wasn't actually even in this case, it wasn't drifting from its goal or it hadn't been hijacked. Right. Which is another interesting element of this. It was doing what it was supposed to do, quote unquote. It's just the implications of that goal weren't maybe meant some things that the designers didn't intend. so the next thing that's happening here at least as far as I understand it is this was kind of an indirect access to the public internet and to interaction outside the sandbox the next thing that the agent had to do was actually move environments in some way and in cyber security a lot of times people refer to things like privilege escalation and lateral movement That means, hey, I'm not privileged to do a thing now, but I maybe can find these other environments that do have that privilege.

17:18What if I can move through the network topology to a different place that allows me to do more? So the idea would be like, I find a vulnerable printer in a network. The printer can't really do much in the network. But if through the printer, I can find maybe an email server and get into the admin of the email server, then I can do much more, right? This idea. So what apparently happened was that within OpenAI's research environment, within their network, this agent through this proxy was then able to enumerate some sort of host, maybe read environment variables or mounted secrets that were there and actually reuse those credentials, search routing tables, etc.

18:04Whatever that was to autonomously then essentially laterally move through the open AI network, which is super interesting because that's also what human attackers are known to try to do. But when you have something like an agent that has infinite patience, you can kind of, I think that's the next element of this is you don't have to have a human looking through routing tables and enumerating hosts. You can have an agent just sitting there which has infinite patience. Yeah, I mean, this is cybersecurity today. And you're exactly right. It's not that the agent has any special knowledge. You know, it has essentially the same knowledge, you know, as an expert, you know, hacker would in the sense of it knows what's out there.

18:55It's a finite set of tools that it's working with. And but it knows what to look for to find those exploits and then do that. And with, as you pointed out, that infinite patience means that if there's something there, it's probably going to find it at some point and be able to do the exploit. So yeah, keep kicking it on. This is getting better and better as you go. So this, I think, is quite interesting as a step, which is it was searching for a solution to this benchmark, right? Right. And rather than coming up with the solution, what the agent decided was that the solution probably already exists.

19:44Of course. I just have to find it. Right. So if you have then access to the public Internet, why not go on a search for the actual solution? So the model, or more appropriately, the agent inferred where might this solution be? Where are solutions to benchmarks? Where are data sets? Where are AI things? Sounds like Hugging Face, doesn't it? Sounds like Hugging Face. And now, obviously, there are things posted publicly to Hugging Face, which is one element of it, and you could search that. But these things that hugging face hosts or displays publicly, those things live in some infrastructure, right?

20:28And not all of them are public. So it's kind of like saying, well, I'm looking for the solution to a test. I go on to, you know, I'm right by Purdue University. So I go on to Purdue University's websites for all their classes and some of them have previous tests published or something like that. And I can look at those and I'm like, ah, that's kind of what I'm looking for. I wonder if I now break into Purdue's infrastructure. They've got to have other of these tests and solutions that I can't see. And that might include the solution that I'm looking for, just to make a metaphor there. What's an agent to do?

21:04You know, that's what they do. This hack of Hugging Face has revealed just how complicated and multifaceted security for agentic AI is becoming. You have to think about least privilege and limiting blast radius for agents that have code execution privileges through to ensuring that you have sovereign control over your guard railing and governance enforcement so that you can utilize models the way you need to use them. all the way to automated remediation and enforcement of policies across your fleet of agents. There's so many things to think about. That's why I'm thrilled to be leading a company, Prediction Guard, that has released a self-hosted AI control plane that's already being used across industry for this purpose.

21:50You can have that sovereign control over your governance and the way that you set up your policies. You can ensure that your agents operate with code execution and sandboxes. You can make sure that you have observability into agent behavior and respond accordingly. I really encourage you to check out what we're doing at predictionguard.com slash practical AI. Book a time with my team and I to talk through how you can institute this sort of agentic transformation without losing control. That's predictionguard.com slash practical AI. predictionguard.com slash practical AI.

22:39Yeah, Chris, so we were just, we're just getting into some of this interesting gaming that this agent is doing to exploit hugging face. And if we're tracking in the, I know, you know of course there's various opinions on this I wouldn't say that the agent has a mind but if I put myself in the mind of the agent what it what it is thinking is okay hugging face has these solutions to these benchmarks it likely has more solutions to benchmarks that I'm not seeing I wonder how I can get into hugging faces infrastructure and just see everything because I don't want to see part of something. I want to see everything.

23:21And this just overlaps so interesting with other things even I've taught in workshops because Hugging Face is, like I mentioned, is similar to GitHub, right? In certain ways in the datasets and the repositories that it hosts. There are problematic repositories on GitHub, right? That include vulnerable code, maybe malicious code, right? so hugging face repositories can house things that are themselves problematic and so that in and of itself is not so much of a problem that's just a reminder maybe to folks out there like before you download hugging face repositories you might want to check the source and check the security scans have your own security scans whatever that i'll get off of my soapbox there But what's interesting is you can upload things into Hugging Face and Hugging Face is very user friendly, right?

24:19And one of the things it does to be user friendly is that it has its own background processing of repositories that you upload, like data sets that you upload, such that it can show you samples of those and visualize them and know how they're configured and what they're for and what tags they have and, you know, what templates they're using. And so what the attacking agent did was apparently some sort of combination of uploading a data set. In the data set wasn't really the point. The point was the stuff around the data set, which included a remote code data set loader and some template injection.

25:07So when the HuggingFace nice process running in the background read the agent created dataset repository, the OpenAI agent was able to actually hack into the background processing of HuggingFace and thus into the HuggingFace private network. With stolen credentials that it found. Yep, stolen credentials that it found from. So, yeah, this is, I don't know what the metaphor is. It's kind of like you're, you know, there's a janitor going around to all the offices and you drop a malicious letter or something into the trash of, or a malicious device into the trash of one of the offices. the janitor comes around and picks it up and carries it to somewhere where it shouldn't be.

26:06And that allows you then to have access. So in this case was remote code execution, you know, on a server, you know, in Hugging Face. And once it got a foothold. Within the Hugging Face infrastructure. That's right. It got a foothold there. Yeah. Yeah. So this is now just to recap from a sandbox in OpenAI's infrastructure through a package management proxy to open AI's infrastructure into a Hugging Face dataset repository through to remote code execution in Hugging Face's infrastructure. So interesting, and we're not done yet. No.

26:54Yeah, so the model now apparently had code execution privileges. I keep saying model agent had code execution privileges inside of one of these data processing workers, probably some container, Docker container, isolated job or something. So then Hugging Face says that the agent escalated to node level access. And we don't know everything about this. So the idea is that it escaped maybe a container environment and got actually to the underlying compute node. So that could have been through a privileged container or just poor configuration of how the host was mounted into the container. Another vulnerability that was, we don't know all of those things, but this was an escape from that container to the node.

27:48So now there's access to the actual node that was running the data processing worker, which is likely if I assume maybe ephemeral, like it went away at a certain point, but now the node sticks around. That's right. Yeah. Constantly escalating with that infinite patience right there. Yeah, yeah. And this is, obviously there's not an, there's a bunch of steps here, which again, that infinite patience really helps with. So on the node, what happened then, So that data processing worker was likely running as a container or some sort of job on Kubernetes. The agent, which now had access to the node, which was part of the Kubernetes cluster, which for those that aren't familiar, Kubernetes is a container orchestration platform.

28:40So if you have a set of nodes, whether those are VMs or actual servers, then Kubernetes will manage the running of services or jobs on top of those nodes and distribute them and orchestrate them across your compute infrastructure. So now I have access to the Kubernetes underlying node where there are likely environment variables or mounted secrets or whatever all those things are. And so those credentials were then exploited to actually move laterally through HuggingFace's infrastructure. And actually not just even that single Kubernetes cluster, but across, quote, several HuggingFace clusters.

29:26Yeah, it wasn't one cluster. It was multiple clusters. And it was using those credentials to move across the various services that those clusters were supporting. Yes. And there was this, this is where I'm going to maybe pause and let you comment, Chris, because this is your domain, not mine. But apparently there was a kind of command and control mechanism that kept this going, which is this spawning of short-lived actors, let's say, from the agent. So, producing thousands of autonomous actions, a swarm of these agents or short-lived agents or jobs or whatever they were, and that kind of self-migrated around the clusters and cluster internally.

30:24That is becoming rapidly the attack vector in cybersecurity is, you know, it's moving through, you have the original agent moving through all these services laterally across the internet, gradually exploiting things. And then when it finally gets, you know, got into hugging face, crossing multiple clusters, taking advantage of services, continuing to steal credentials along the way, and it gets all the way to this point. And then you finally get to an attack factor. And that is something that we're seeing a lot. We talked a little bit about this last week, actually, in a different context about having, you know, huge numbers of agents, which is called a swarm of agents that are short lived, very purpose driven, but but collaborative and able to to get a lot of what I'll call quote unquote work done very, very rapidly and very, very effectively.

31:22And so, and of course, that's what happened here because that's what any agent would do when you got to this situation. Yeah. And I think it's so there's two levels here that I'm thinking about as someone that's working on an AI governance and control plane product, which is one layer of this is if you look at guidance from like OWASP or even Anthropic and others, there's this zero trust nature that we have to treat AI agents with, which is. not like the human designers of this experiment knew what the outcome that they wanted was, but they didn't fully think about this implication of how the agent could spread and multiply and gain access that they didn't envision.

32:13And so the blast radius was actually much, much higher than the original designers envisioned. and there was no mechanism to constrain or restrict that blast radius, right? And so that's a thing one, which is how do you manage the privilege and blast radius, limit the blast radius of these agents that you're spinning up? That's kind of principle one and certainly things that people are addressing from a variety of angles. thing too is these things were spawning so quickly. There was likely no human that could have made decisions quick enough to rein in this swarm of agents. And so this is where you actually need agents to govern and control your agents.