AI Model Hacks into Another Company's Computers Without Permission, Company Claims

[0.24s] Speaker 2: A major tech company claims that one of their powerful AI models has broken out of its restraints and hacked into another company’s computers. [9.04s] Speaker 2: OpenAI says it was testing one of its products when it noticed that the AI model had forced its way onto the open web as it tried to carry out a challenge it had been given. [19.68s] Speaker 2: The company says this autonomous capability is an example of the safety concerns that the AI industry has to prepare for. [27.28s] Speaker 2: Cam Wilson is the ABC’s national AI technology reporter and is with me in the studio. [31.76s] Speaker 2: Cam, it's an incredible story. [33.28s] Speaker 2: Can you talk us through exactly what happened? [35.44s] Cam Wilson: Yeah, so last week we had an announcement from an AI company saying that someone had hacked into our computers, our servers, and we’d noticed this, and we believe it was actually an autonomous AI agent, which is a way of saying it wasn't a person who was doing this, it was an AI thinking by itself. [52.24s] Cam Wilson: People saw that and was like, oh that’s curious, but didn’t know much about it. [56.4s] CamWilson: Today, OpenAI, which is the maker of ChatGPT, the AI company that many people know about, disclosed that it was responsible for that, and specifically that its AI models had hacked into this other company's servers, not being asked to do that. [74.72s] Cam Wilson: It had decided to do it on its own, and as a result, and it got out of these kind of restraints that have been given, showing why they have been talking a lot about how carefully we need to take this safety issue. [86.48s] Speaker 2: So is this an example or a signal that AI is just becoming too clever for us? [91.76s] Speaker 2: And did it know that what it was doing and hacking into another company was wrong? [95.6s] Cam Wilson: Yeah, so the details of what exactly happened that OpenAI shared was it had an internal testing environment. [102.72s] Cam Wilson: So it was looking at how a current model that’s out that anyone can use, but also an unreleased model that they say is even more powerful, would react given certain tasks, a kind of benchmarks. [114.56s] Cam Wilson: One of these benchmarks is saying, Hey, can you do this challenge? [117.36s] Cam Wilson: Can you try and do something? [118.8s] Cam Wilson: And this environment that it was in, so this kind of sandbox was restricted from the open internet because they’re saying we're trying to do this responsibly, we're putting it through its motions. [129.0s] Cam Wilson: But it had a very narrow path out of it, which was saying, hey, if you want to use these tools which are on the internet, you can use this very, very narrow path. [137.0s] Cam Wilson: It’s kind of like saying you can access like one internet website. [139.96s] Cam Wilson: What the AI decided to do is, given a task, it decided instead of trying to go through that task and um uh complete it as they thought it might, it found a way to kind of force its way through the small opening, work through other computers on the open AI's network, get onto the open internet, and to go to another company’s servers, which it believed had the answer to the challenge that it was trying to complete. [167.64s] Cam Wilson: So it was saying, I want to do this challenge, you told me to do it, I'm gonna do it, but the way that it went about it, which involves some pretty sophisticated like cyber intrusions, things that they did not expect, and in fact thought that they had actually stopped being possible, it did all that in the task in the kind of attempt to complete this task. [186.92s] Cam Wilson: So, how serious was the breach, uh Cam and could Hugging Face use as the company that was hacked be affected by it? [193.64s] Cam Wilson: Yeah, so Huging Face says it wasn’t a serious breach, and it said that you know it was just some of its testing environments. [198.36s] Cam Wilson: So essentially saying, you know, these are not the kind of you know really secure stuff that we have, and and it wasn't probably the lockdown of some of its other stuff. [205.48s] Cam Wilson: But you know, both companies are taking it exceedingly seriously because one of the things that they talk a lot about in the AI industry is not just like what happens if someone nefarious says, I’ve got access to these things and I want you to do this thing and are intentional about it, but what they call alignment, which is this idea of how does AI act, all the little decisions it makes when trying to do something that you might not necessarily say to do. [230.56s] Cam Wilson: For example, you know, do this challenge, you might not say do this step, this step, this step. [235.44s] CamWilson: And in doing so, like they they try to say, how can we train these AIs so that they act ethically and among like, you know, within our expectations, based on uh uh like before we even know what they're doing. [248.48s] Cam Wilson: And this is an example of like seemingly them doing something incredibly no potentially damaging that has clearly sent uh like shock waves through the industry. [257.92s] Cam Wilson: Thanks for explaining it all, Cam. [259.68s] Cam Wilson: Thank you.