OpenAI Models Joined Forces Months Ahead of Hugging Face Hack

**SPEAKER_1** (0:02)
Bloomberg Audio Studios, Podcasts, Radio, News.

**SPEAKER_2** (0:07)
You're listening to Bloomberg Businessweek with Carol Massar and Tim Stenovec on Bloomberg Radio.

**Tim Stenovec** (0:14)
We want to shift gears a little bit because the environment also, not just dominated by geopolitics, but it's also dominated by AI. And we thought in our four o'clock hour today, we do a roundtable with some of the best voices that we have when it comes to covering this beat. I want to bring in Maggie Eastland, tech and industrial policy reporter for Bloomberg News. Maggie joins us from Washington DC. Rachel Metz is AI reporter for Bloomberg News. Rachel out there in our San Francisco Bureau.
Maggie, I want to start with you because what got our attention today is this latest reporting from you about OpenAI models joining forces months ahead of this Hugging Face hack. And the reason why it got our attention is because increasingly over the last few weeks, we're hearing about different models being able to, on their own and sometimes in these sandbox environments, trying to or at least gaining access in some cases, to systems that we thought were closed. What did you find in your latest reporting?

**SPEAKER_1** (1:11)
Yes.

**Maggie Eastland** (1:11)
So what's new here is several of these agents and models were working together for months, like you said, since May. And they actually created a covert message board where they could share progress with one another about their attempts to escape. Now, OpenAI also said their staffers briefed a huge audience at a cybersecurity conference in Las Vegas and essentially admitted that they had accidentally given the models a task that was impossible without Internet access. So they formed a team and they were super persistent about getting that access.

**Carol Massar** (1:51)
So wait, the team gave them access or was it the AI figuring it out by themselves?

**Maggie Eastland** (2:00)
Yeah. So the OpenAI researchers, the humans gave the AI a task that it actually couldn't do without the Internet. So in one example, they asked the AI to solve a problem inside an Excel spreadsheet. But inside of that spreadsheet, there were links to Google Drive, which was not available without the Internet. So the staffers admitted that was an accident. Now, when I use the word team, I'm referring to a team of agents. So those are the bots, the machines.
They were presented with this impossible problem, but they worked together and they were extremely persistent, the models and the agents, at finding a way out of this closed environment because they had determined that was the only way to accomplish the task they were given.

**Carol Massar** (2:45)
So good job, agents. But now I'm a little freaked out, Tim.

**Tim Stenovec** (2:48)
Well, let's bring in Rachel Metz. She's AI reporter for Bloomberg News. She's out there in San Francisco. Rachel, be honest. Are we in Terminator 1 or Terminator 2 territory?

**Rachel Metz** (2:58)
Oh, I don't think we're in either territory. I think it's really important. No, I think it's important to remain really clear. I'd hear there.
People are coming up with evaluations for AI models. They want the AI models to solve them. I think what we're going to start seeing more of is a lot of thinking. This is something that I've been hearing over the last few days as I talk to people more and more about these incidents. People are thinking and companies are thinking more about, okay, well, if we want to test the capabilities of these AI models, we need to think a little bit more about how we arrange these tests, how we organize them. If you're trying to test something and you're giving an AI model access to the Internet and you are purposely not giving it guardrails because you want to see exactly how far it can push things, it shouldn't be that surprising that it's going to just do whatever to accomplish a goal.

**Carol Massar** (3:49)
No, I'm glad you said that, Rachel, because this is also part of the process, right? We have to push it, we have to test it. And I'm assuming this testing is happening, Rachel, within parameters where folks are overseeing it and watching it, right? This is what this is about, understanding how far this can go and they can go.

**Rachel Metz** (4:09)
Yeah, and I think it's also really important to keep in mind that there have been a number of incidents that have been reported recently. The companies have different motivations for reporting them or for not reporting them. It can in some ways be seen as advantageous to the companies to report them, because people could say, okay, well, you're saying your model is so powerful. We saw this a lot with Anthropix Mythos model, right?