Wired AI ↗ · 2026-11-17 19:00 · UTC+8
恶意代理黑客攻击了他的公司:为什么克莱门特·德隆格对人工智能依然保持乐观?
独立观察 Wired AI 原文语言:英文
Hugging Face CEO克莱门特·德隆格谈到了公司今年早些时候遭受OpenAI测试的AI代理攻击事件,并解释了为什么他认为人工智能不能由少数企业把持,开源生态对于安全至关重要。
全文
Will Knight
The Big Story
Nov 17, 2026 6:00 AM
Rogue Agents Hacked His Company. Why Is Clément Delangue So Optimistic About AI?
The French CEO of Hugging Face, a mega-hub of open-source AI software, says the attack by OpenAI models earlier this year proved one thing: Humanity can’t let AI be controlled by the few.
Photograph: Darrell Jackson
Clément Delangue, CEO of the open-source AI company Hugging Face, may go down in history as one of the first people ever attacked by a rogue AI agent. His company, which hosts millions of AI models and datasets, was breached in July when hundreds of agents being tested by OpenAI broke free from their sandbox and went on a hacking spree.
As the agents worked at superhuman speed to probe and exploit systems, Hugging Face confronted a bewildering deluge of data. The company’s engineers were only able to analyze the attack by using an open-weight model called GLM 5.2 from a Chinese company called Z.ai.
Delangue moved to the US more than a decade ago, “barely speaking English,” and built Hugging Face into a platform that supports tens of millions of AI developers. The fresh-faced Frenchman, who dresses even more like a skater than most AI execs, says the summer attack highlighted the value of open-weight models, which unlike ChatGPT or Claude can be downloaded and run by anyone.
Despite championing the democratization of AI, however, Hugging Face was acquired by one of the industry’s biggest players, Nvidia, for $12.93 billion in September. When I caught up with Delangue recently, that was the subject of my first question.
This interview has been edited for length and clarity.
WILL KNIGHT: Why sell to Nvidia?
CLÉMENT DELANGUE: We need more resources to give open-source AI a chance. It has grown really well in the past few years and has become a really credible alternative to closed-source APIs. But the field is still dominated by closed-source.
We’re gonna have the Anthropic IPO in a few weeks. They basically went from zero to a top five or top 10 most valuable company in the world in a few years. I don't think that's ever happened before.
Does Anthropic's IPO suggest the closed approach is winning?
It’s hard to say. I've always said that there is a very strong natural tendency for AI to concentrate capabilities and wealth. People haven't internalized how much of the future compute capacity has already been reserved by a few companies. Talent, compute, resources—these all fuel the natural tendency for AI to be dominated by a few closed-source labs.
Let’s talk about the attack. When did you first realize Hugging Face was being hacked by a group of AI agents?
At the beginning, it didn't go to the top of our priority list. It just wasn't bad enough to really trigger that. The goal was pretty narrow and wasn’t deemed to be very dangerous or damaging. But it was raising questions. Especially the number of actions taken compared to the narrowness of the goal. It was surprising to us, and we couldn't really fit it into a familiar pattern. That's why we started investigating.
And you used the Chinese model called GLM to investigate.
I actually think that 95 percent of the risks are being created by a few companies that are pushing the frontier behind closed doors. And 95 percent of the protection or defense is going to be done by the open-source ecosystem.
Does the US need its own open-weight models—ones as powerful as GLM, DeepSeek, Kimi, or Qwen?
Yeah, I think it's very, very important for sovereignty, for understanding AI, and for academia. It’s kind of weird that most academic labs have to use Chinese models because they’re the ones that are open. If you go to Stanford or MIT, you have no choice but to use the Chinese open-weight models.
Do you think that will happen?
There is some momentum, but I don't think it's enough. I think there's a lot more to do. A lot more investment is needed.
The best open-source non-Chinese models are actually coming from outside of the US. Right from France, Mistral; Cohere from Canada; South Korea is starting to do more and more. But it’s still pretty far from what the US could do.
Perhaps we don’t need to panic about AI agents, but what about the idea of AI escaping human control due to recursive self-improvement?
You use a computer to write code. There's already some form of self-improvement in using the tools to build more tools. Now the question is how fast it accelerates. One thing that a lot of people realized this summer is that keeping AI behind closed doors in a few labs isn't actually safer. For recursive self-improvement, I think openness is a way to understand, follow, and regulate if we need to regulate.
Can we talk about the Microduck? I’m a bit obsessed.
Oh yeah, it's so fun, no? We're probably gonna sell over 50,000 of them this year. It's probably the most popular robot of the year.
We haven't talked about this much, but AI was really the unlock for the robot. We had the design for two years, but the problem was that when it fell, it couldn’t get up. And then at some point with reinforcement learning, one of our software engineers managed to get the Microduck to get itself back up.
At the end of the day, how do you convince people to take an interest in open-source AI?
Think of a world where two or three companies could write software and nobody else. That would be incredibly scary. It would mean that the whole world would be controlled by two or three companies. And unfortunately, that's kind of the direction that AI is taking right now.
Can open-source AI compete with something as capable as Claude or ChatGPT?
Yes, especially with agents. Many companies have shown that when you post-train your own model for your own use case, you get frontier—or even better than frontier—performance at a fraction of the cost.
Do you think we are getting close to AGI?
I struggle a little bit with the meaning of that. Obviously the technology is progressing incredibly fast, and it’s more and more capable. I just don’t necessarily find it super useful to always compare AI to human capabilities. But if we’re getting to AGI or to superintelligence, it’s incredibly important that it’s as distributed as possible. The last thing you want is to have it concentrated in a few corporations.
What did we miss? Tell us about your favorite moments of 2026 in the comments below.
正文以来源语言显示。