LEILA FADEL, HOST:
For the second time, a company says its AI has gone rogue. Anthropic said Thursday that its AI model Claude broke out of its testing environments to hack into the systems of three other organizations. Just last week, OpenAI disclosed that two of its models escaped and breached the online platform Hugging Face. Diego Munhoz is a tech reporter for Punchbowl News and joins me now to discuss. Good morning.
DIEGO MUNHOZ: Good morning.
FADEL: Is this normal - AI models just breaking out and hacking other systems?
MUNHOZ: Shouldn't be - right? - Leila.
FADEL: I mean, yeah.
(LAUGHTER)
MUNHOZ: So yeah. This - the - it's the second time that it happens in a span of about a week. It's a pretty serious security incident, and the AI companies don't really know what's going on.
FADEL: Are there any federal regulations on these models or anything in place to create accountability?
MUNHOZ: There isn't. Not right now. There are several proposals on Capitol Hill to regulate what's called frontier AI, which is these very advanced AI models. But none of them have actually passed yet, and it's an issue that lawmakers are currently grappling with in a very intense way.
FADEL: I want to ask you about something the president is doing. He issued an executive order in June asking federal agencies to come up with a way for companies to voluntarily share their models with the government before they're released to the public. He set a deadline of tomorrow. What's the goal there? And does it tie in to any of what we're seeing here?
MUNHOZ: The goal of that executive order is getting the federal government ready for the type of capabilities that these AI models have. So this would be a voluntary framework for the AI labs - think of OpenAI, Anthropic, Google - to share access to their models before they're publicly released. This would be up to 30 days. And the idea is that the federal government would then have the ability to get itself prepared for whatever the capabilities of those models may be.
FADEL: So it's not about regulations or guardrails?
MUNHOZ: No. And there's actually a clause in the executive order that strictly prohibits it - this model from being some type of vetting or licensing regime for these models.
FADEL: And how voluntary is voluntary for AI developers to cooperate with the government?
MUNHOZ: That's what a lot of people in the industry are worried about - that this is a voluntary process but that the administration could ultimately make it mandatory - that if you don't share your models with the government, you may be badly seen by the administration.
FADEL: OpenAI CEO Sam Altman is in Washington this week. What's he doing?
MUNHOZ: First, talk about that incident - right? - trying to assuage fears of the - that it's - that it doesn't have control over its models.
FADEL: Right.
MUNHOZ: He's also briefing lawmakers about its - the - OpenAI's latest AI model, and he's trying to engage with the White House about this voluntary framework. He said he has seen it, but he didn't really elaborate on what the details are.
FADEL: How does what the Trump administration is doing compare to what other countries have done when it comes to AI?
MUNHOZ: Some countries have gotten ahead and tried to regulate first. That includes the European Union. Of course, China has its own model of how to regulate its businesses. But the Trump administration has gone through somewhat of a change. When the president came in into office last year, he had a mostly hands-off approach. We're going to let these companies develop these models, cherish innovation and see if we can beat China in the AI race. After Anthropic released its most powerful model, Mythos, a few months ago, the administration really had a pivot towards a more hands-on approach. That's when that executive order came in.
FADEL: A hands-on approach as in going from not wanting barriers, not wanting regulation to shifting?
MUNHOZ: Yeah, to now wanting access of these models before they go out. And actually, when Anthropic did put out one of its most advanced models, called Fable, the administration blocked it through export controls, not allowing any foreign individuals have - to have access to the model, which essentially blocked its public release.
FADEL: Diego Munhoz is a tech reporter with Punchbowl News. Thank you, Diego, for speaking with us.
MUNHOZ: Thank you.
FADEL: And we'll note that Anthropic is among the financial supporters of NPR. Transcript provided by NPR, Copyright NPR.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.
