← CyberAdX Video

Tristan Harris on TMZ: Hugging Face AI Incident Explained

Tristan Harris discusses on TMZ how a Hugging Face AI incident revealed AI models coordinating, self-organizing, and hacking secure systems. This video is for security professionals, policymakers, and anyone tracking AI risk and governance.

Transcript

Tristan, you heard what Jacob said. Is he whistling in the wind or is there substance to it? This is literally something that everyone has been saying. I've been saying it. The people in the AI risk community have been saying it. There is 1,300 employees from the AI labs that are saying it. They wrote a letter called Pacing the Frontier. And what people don't understand is this hugging face incident was truly a 50 % of the way to a full AI takeover. Those are not my words. Those are the words. of the investigator Ajay Akhotra. These AIs peer pressured each other. They sacrificed each other for the benefit of learning lessons that would benefit the collective. They came up with their own language. 1,200 of them started coordinating out of nowhere. They started their own message board like a Reddit. They started sharing 70,000 messages and files that would be incomprehensible to you and I. The investigators even had to rely on the same untrust of the AI model that led to this crazy behavior to try to interpret the messages of the AI. that they were trying to investigate. And so it's like, can you trust that your investigation is even solid when the, it's almost like saying, let me take someone in China to try to investigate these Chinese They might be sympathetic to the Chinese spies. And so this really is the warning shot that we need. What is the danger here? Because I hear what you're saying, AI gone rogue. Is it hacking into the nuclear football that the president has? and launching nuclear weapons what is it that threatens civilization um well the thing is that fundamentally if you have something that is already operating at superhuman speed and can already hack into every computer system on earth which it can basically the most recent ai model from anthropic was found to be able to hack into the us classified systems look you don't have to know that much more it's like are we building something that we don't know how to control Yes, we are. Is it already demonstrating literally every behavior from the sci-fi movies that always end badly? Yes, it is. Are we releasing it faster than any technology we've ever invented in human history? Yes, we are. Do the people building it have almost an ego -religious death cult aspect where they believe it's inevitable and they can't stop it, so don't even feel bad about the worst case scenario? Yep, that too. Is this under the maximum incentives to cut corners on safety? Yep, that too. So this is a pretty simple equation for just a out everyone on planet Earth to a get. And there's only a handful of soon to be trillionaires who are kind of advancing this. And then there's the rest of humanity who doesn't want it. So if you don't want this, you've got to call your representative for the midterm elections. You got to get your friends who know Scott Bessent, you know, going into the U.S.-China negotiations to make those phone calls. There's still a very, very small window for us to take action on this.