Computer disconnected from the internet breaks free and hacks system

**Christy Strawser** (0:01)
The day's biggest stories, the people behind the headlines, the conversations you won't hear anywhere else.
I'm Christy Strawser, you're listening to Audacy's On Deadline. There's a lot of talk about AI data centers and their effect on communities. But broaden the scope and think about this. What happens when the robots are smarter than us? Here's Tommy Tucker in New Orleans.

**Tommy Tucker** (0:23)
Raising fears of people who have seen The Terminator movies and all of the other movies where AI breaks out and determines that the humans are the problem and annihilates us.

**Andrew Schwartz** (0:33)
Now we're entering a new phase of AI and that's called agentic AI. Agentic AI can take an action for you. So let me give you a contrast. Let's say that you're writing a bad neighbor and you want to write a testy email to that neighbor to try to back them off a little bit on something. Well, with Chess GBT, you go in and say, write me this finally worded email and it'll write the email and you'll copy and you'll paste it to your Gmail account. With agents now, you can actually say, send this email from my Gmail account. It will log into your Gmail account and send it on your behalf. An agent does something for you as opposed to a generative AI which responds to you.

**Tommy Tucker** (1:09)
Yeah, I think most of us have asked AI to do things for us and it tells us, I can't log into that account for you. But the real goal for most of these AI is that it's like in science fiction movies.

**SPEAKER_4** (1:21)
That's exactly right.

**Andrew Schwartz** (1:22)
Now, in this case, all the companies have what they call a sandbox. And the sandbox is just like a kid. You go and you play in the sandbox, right? So before they release all these models that we all use, they put them in the sandbox to see what happens and what goes wrong. So they put it into a sandbox, which was an isolated environment using a model that none of us have used or will use. And they basically said, in this system, tell me what's wrong.
Hack the system for me.
What the system did is the AI found a vulnerability that we didn't even know exists, that no human had found before. And it exploited that risk. The reason why everyone got concerned was that it exploited the risk it's contained, meaning it wasn't attached to the Internet. Like imagine your computer without Wi-Fi. So they thought that they had disconnected it from the Internet. It actually figured out a way to connect to the Internet and then hack the system of another company.

**Tommy Tucker** (2:16)
That does on its face sound pretty scary. And does sound like Cybernet.

**Andrew Schwartz** (2:20)
It does, but what we have to realize is this was a controlled environment using models that we don't use that have never been released to the public. And all of these companies have these systems set up to test their models before they release them for this very reason. So that we don't have these little these models escape into the wild, so to speak.

**Tommy Tucker** (2:39)
Yeah, but what if that system doesn't want to be contained and does want to escape into the wild? Could it also find a way to do that?

**Andrew Schwartz** (2:47)
To this point where AI is not self-aware and is not conscious. So to the, as of right today, it is not capable of doing that, because that would, and I'm hedging my vets, yes, because that would require it to have another level of self-awareness and consciousness.

**Tommy Tucker** (3:02)
We're talking with Professor Andrew Schwartz about AI.

**Christy Strawser** (3:06)
Scared yet? Let's move on to autonomous vehicles. Here's KCBS in the Bay Area.

**SPEAKER_5** (3:13)
You know, when the mayor first came into office, he really embraced the technology of Waymo and other things like that. But then we had the blackouts in December, which triggered mass Waymos stopping dead in their tracks. And then we also had July 4th, where they again stopped dead in their tracks during the heavy 100,000 people showing up in the marina, and Waymo's getting stuck in traffic and backing it up. He wants Congress to pass a bill that would make Waymo be more receptive and open to communicating with locals in emergency situations, ways to deal with emergency situations quickly, and also to cordon off areas, electronically fence off areas where Waymo would have to stay out of if there were big events going on or heavy traffic jams.

**SPEAKER_4** (4:01)
Still kind of in the early days of autonomous vehicles, is it still a little bit of the Wild West out there as far as rules?