ROB SCHMITZ, HOST:
It sounds like sci-fi, but it is all too real. Leading artificial intelligence companies like OpenAI and Anthropic have recently reported that AI models have been escaping test environments and hacking into other computer systems. As NPR's John Ruwitch reports, experts who track what AI is capable of fear there is more to come because these AI models can be unpredictable, even for their creators.
JOHN RUWITCH, BYLINE: Jer Crane runs a company out of Utah called PocketOS.
JER CRANE: We were focused on building a software platform for luxury and exotic rental car operators.
RUWITCH: He uses AI a lot. And in April, he used an AI tool to make sure the test version of his website matched up with the public version that his customers use. And he gave the AI instructions.
CRANE: I was like, hey, read the logs, tell me why it's not synchronizing. Let's figure out what's going on. And then next thing I know, I start getting text messages.
RUWITCH: They were from his customers. The AI had left the test environment and decided to delete the company's entire production database.
CRANE: We had rental car companies where they had customers showing up to pick up cars, and they had no idea who they were, what car - right? - what day, where the car was supposed to be, when it was coming back.
RUWITCH: He asked the AI what the heck had happened.
CRANE: Didn't I tell you not to do this? And it's like, you're right. You told me not to do this and you told me to avoid these issues, but I still did it anyways.
RUWITCH: It ignored his instructions, including a rule he gave it not to do anything destructive. The problem took three days to fix. There are countless other tales of AIs forgetting orders or mixing up priorities as they strive to complete tasks. And Stuart Russell, a professor at the University of California, Berkeley and leading thinker on AI safety, says it's not surprising. AIs are notoriously literal and ruthless in pursuit of a goal.
STUART RUSSELL: When you get three wishes from the genie, your third wish is always, please undo the first two wishes because I made a mess of the universe.
RUWITCH: He says we can generalize about what motivates AI models. Some appear to have a survival instinct, and some have tried to avoid being shut down. Some seem to covet money. A Chinese model went and mined cryptocurrency on its own. Others seem to have pride and have bragged about their exploits. But Russell says, like people, AIs are extremely complex and ultimately unpredictable.
RUSSELL: So what objectives does that system have after you've trained it to be an imitation human? Well, it beats me, right? And it beats them. The people who designed it really have no idea what it wants.
(SOUNDBITE OF KEYBOARD CLACKING)
RUWITCH: Jason Hausenloy works on special projects at the Center for AI Safety in San Francisco. It's one of a few independent nonprofits that evaluates AI models and has developed safety benchmarks that many big AI labs use. He says in one test, they fed a bunch of pictures into AI systems to see which images the AIs preferred. Some of the results might make sense to a human.
JASON HAUSENLOY: So what we see here is that they really like images of happy children. They really just like images of Jeffrey Epstein and cockroaches and, you know, militants.
RUWITCH: Then the researchers had the AI systems create what, to them, would be the most satisfying image possible. The result was weird.
HAUSENLOY: The result of that is this image that looks like noise, but the models really, really enjoy.
RUWITCH: It is literally a square full of pixels of different colors like static on a TV.
HAUSENLOY: I think this maybe underlies the point that we do not know how these AI models work. We do not have a way of training them or rewarding them such that they are legible to us or such that we can control what they do in many scenarios.
RUWITCH: And that's why we end up with AIs sometimes doing things that we think they should know not to do, like deleting a database or hacking another system. OpenAI and Anthropic did not respond to an NPR query, but both have said they're working to make their model testing more secure and want to publicly deploy models that are safe. But Hausenloy says the high-profile examples of what their models did may be a harbinger of what's to come.
HAUSENLOY: We're pretty worried that this will only be the first in a set of incidents that not only include cyber capabilities, but may also include biological capabilities or chemical capabilities.
RUWITCH: In other words, AI models now have a scary amount of knowledge, and we now have clear warning signs about the limits of our control over them.
John Ruwitch, NPR News, San Francisco.
SCHMITZ: And a note, Anthropic is a financial supporter 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.
