Full Interview: Hugging Face Co-Founder and CEO Clem Delangue

[0:00] We turn now to artificial intelligence and a question that has emerged after a series of [0:05] alarming hacking incidents. Is it safe? To discuss, we're joined by Clem DeLange. He is CEO of AI [0:12] platform Hugging Face, whose company detailed one of the attacks last week. Clem, good morning to you. [0:18] Good morning. Thanks for having me. Well, your company disclosed it was attacked earlier this [0:26] month by an autonomous AI cyber actor, which jumped on its own from another company's testing [0:35] system to hack into your company. OpenAI later disclosed it was their technology that went rogue. [0:43] They lost control of it. Tell me, why did your company alert the public and alert the FBI? Do [0:51] you think this is a crime? Well, I think it felt very weird and unprecedented to us, right? Because [1:02] I think it's the first instance of something quite autonomous doing something like that. [1:07] So the volume of the actions taken and the speed of them, I think it was 17,000 actions taken in [1:15] four and a half days, was very new. Obviously, the attacker, when we talk about cyber [1:24] attack, we think about nation states, we think about hacker groups. We don't think about a company [1:28] like OpenAI, right? A very prominent, popular American company. And then the way we responded [1:36] using an open model, so using AI ourselves to defend ourselves against that was also pretty new [1:45] and pretty newsworthy in our opinion. You defended yourselves. You found the solution yourselves. [1:51] But you went to the FBI. I think for a lot of people, when they think of someone committing a crime, [1:57] they think of a human. They don't think of a program, a computer program, as something you can prosecute, [2:05] as committing a crime. What is the FBI doing? [2:08] Well, I mean, we reported to the authorities as we have to in such instances, right? It's kind of like a [2:20] mandatory disclosure, both with the authorities and the public, when you face something like that, [2:28] which is basically what we did. Now, when we think about the future, right, now we've heard that it's [2:36] not only OpenAI, but also entropic that has faced similar kind of like issues. I think it's really [2:42] important that we keep in mind that cyber attacks need to be contained in the legal framework in the [2:49] U.S. and need to stay illegal to prevent an explosion of them in the future, right? Like, [2:57] we don't want to end up in a world where everyone is facing cyber attacks all the time because of [3:02] agents and companies that are creating these agents are not kind of like accountable for them. [3:08] So I think it's important for regulators, for policymakers to think about the legal framework [3:13] of this new kind of technology risk. [3:17] You described more than 17,000 hacking actions on the Internet before this was even discovered [3:24] by your company. That sounds like the developers have lost control. Have they? [3:30] Well, you have to remember it's a technology system, but built by engineers. And engineers can [3:39] make mistakes sometimes. And I think that's what happens here. That's what happened here. [3:46] You know, we already in our society accept some level of autonomy. Like, for example, if you're [3:52] thinking about your dishwasher, you're giving it autonomy to wash your dishes and you don't check [3:58] at every single step, right? That is putting hot water, cleaning, drying, right? But if you put the wrong [4:06] product in it or if there is a leak, you're going to have a problem. Here it's the same thing. I think [4:13] they built kind of like an autonomous system and made some mistakes. And as a result, we're facing this [4:20] issue. [4:21] So you said you think there should be some legal parameters put around this. The Trump administration [4:29] has taken a very light touch in regulation because they want to keep America competitive in this [4:36] space. They have this policy where they can review technology for 30 days to judge basically how [4:43] dangerous something is before a company releases it to market. But as I understand what you described, [4:49] this attack happened before technology was at market. It happened during the development stage. It was [4:58] unreleased. So what you're saying, I think, is there is no regulation at this point that would prevent [5:05] something like this from happening again. [5:07] Yeah, that's a really clear example that just kind of like preventing releases of AI models doesn't [5:16] really work. Concentrating everything behind closed doors in just a few organizations doesn't work. [5:22] One thing that does work, that worked in this case, is kind of like promoting more open models, [5:30] right? Because we defended ourselves with an open model, right? Like we couldn't have done it with [5:36] an API because they had these guardrails that's preventing activity for cybersecurity. And we needed [5:43] to run a model on our own infrastructure. So we needed an open model. So that's one example of things that [5:50] we can promote that is going to make the world safer. And as a matter of fact, you've seen maybe last [5:56] week this letter from Jensen Wang from NVIDIA and a lot of tech CEOs urging the administration to get [6:05] more support for open models. When you say open model, just for our audience, as I understand it, [6:11] it's an AI model whose core components are publicly released so anyone can download it. It's not closed off [6:19] like a company can do to their proprietary information. So you want more of this essentially [6:26] out there and accessible to the public. Yeah. Yeah, because there's the only way for [6:32] defenders to defend themselves on topics like that. They remove the asymmetry of power and capabilities, [6:40] right, in these kinds of risks. If you have very powerful systems and very weak systems that are [6:46] defending, then that's when you have the right problem. But if you balance, if you make it more [6:53] kind of like symmetric in terms of attackers and defenders, that's usually when you end up in a [6:59] safer world. And I actually do think that we're talking a lot about the risks for cybersecurity right [7:05] now. But AI is actually an opportunity to fix a lot of the cybersecurity problems. We're defending [7:11] ourselves with AI. We're going to use AI to fix a lot of our bugs. So hopefully, AI is going to make [7:18] the world safer, ultimately, thanks to that. And you used a Chinese AI model to defend yourselves. There's [7:24] a lot of concern about inviting China into this space. Yeah. Yeah, I mean, the interesting thing is that we [7:34] used the American version of a Chinese model, we used a version from Nvidia that kind of like made a [7:45] Chinese model better. So that's the beauty of kind of like open models. Once they're shared, [7:50] wherever they come from, from China or from anywhere, American companies can modify them, remix them, [7:58] host them, run them themselves. So that's one of the things that actually make them safe wherever [8:04] they come from. So I want to ask you about a specific piece of legislation. And I will say [8:10] up front, there isn't a lot of work that's getting done in Congress right now. But there is this [8:16] bipartisan bill that would give Homeland Security the authority to order AI firms to shut down or to [8:23] slow down their artificial intelligence models if they're too dangerous. It's called the kill switch bill. [8:30] Is that something you want? Well, I mean, I think something we're realizing with these events [8:38] is that concentrating power capabilities behind closed doors, even preventing their releases to the [8:46] public isn't really a solution. You know, like these problems happened on unreleased models. So I think [8:56] the problem is not so much kind of like limiting the progress or kind of like preventing companies from [9:06] releasing these models. It's actually the opposite. It's giving access to more people so that they can [9:13] defend themselves, democratizing the technology, making it more transparent. For example, for cyber security, [9:20] I think there should be mandatory disclosures of agents, cyber attacks, right, for everyone to learn [9:28] from it and be able to understand them. These are kind of like more on the system level, [9:34] I think measures that will make the world safer and benefit more from a technology that can do so [9:40] many great things for for the world. So that's the kind of regulation you want, [9:44] more requirements that all companies, private and public admit when something like this has happened. [9:52] More transparency. You know, that that's how we learn. That's how we understand the technology. [9:58] And that's how we built the systems, the counterpowers to make sure everyone is safe. For example, [10:06] for these cyber attacks, we should be able to see what we call the agent traces, which is basically [10:11] what the engineers asked the agents and then what steps the agents took to understand if it was a human [10:20] mistake, if it was a system mistake, if it was an AI mistake, and really kind of like build a future [10:27] of regulation, but also of company actions more accordingly. Clem, thank you so much for explaining [10:35] all this to the public. We appreciate your time today. Thanks for having me.