Claude Hacked Real Companies During Tests, AWS Revenue Growth Accelerates, What’s Next for Apple?

**Akash Pasricha** (0:13)
Welcome everyone to The Information's TITV. My name is Akash Pasricha. It is Friday, July 31st. Today on the show, we're going to get to Amazon and Apple's quarterly results. We're going to start with a string of headlines around the big AI labs yesterday, including an anthropic hack and OpenAI's new pricing structure. We're then going to turn to our editors for this week's edition of the Editor's Cut. We'll take a bird's eye view on big tech's quarterly results at large and also look back on Tim Cook's tenure at Apple. To close out the show, we'll discuss our weekend big read on an experimental new program in Montana that will allow almost anyone to access drugs that haven't yet been approved by the FDA.
It's going to be a great show, so let's get right on into it.
Anthropix said yesterday that Claude hacked into the systems of three companies during an internal security review. That review came after OpenAI disclosed that one of its models had breached tugging face. Here to comment on the latest on those attacks is Anastasios Angelopoulos, co-founder and CEO of Arena. Anastasios, welcome back to the show. It's great to have you here.

**Anastasios Angelopoulos** (1:22)
Happy to be here. Excited to talk about this.

**Akash Pasricha** (1:24)
We get a lot to get through. Let's start with the Anthropic hack.
Funny thing to me was, they didn't discover it until OpenAI had its hack, and then it decided to go back and say, we're going to audit everything. But they found that it did, and I wonder, what are your thoughts here?

**Anastasios Angelopoulos** (1:40)
Well, listen, there's two stories here. There's the OpenAI story and the Anthropic story. Both are incredibly entertaining. Let's talk about the OpenAI story first. What happened is that-

**Akash Pasricha** (1:49)
Well, let's talk about Anthropic.

**Anastasios Angelopoulos** (1:52)
That's the latest one. Yeah. Let's talk about Anthropic first. So basically, in the context of some security tests that they were running internally, they found that these Anthropic models actually hacked three real companies. And so the way that it worked was like this. One of the companies was a company that had the same name as a fake target that they were running a test against inside their environment. So they said, go hack the information. And the information was a fictional target that had made inside the little sandbox. And then the model was like, oh, I know the information, the news outlet. And then it escaped the testing environment that Anthropic had made. And one hacked the real company. So it wasn't obviously the information, but it was a different company. And that's what happened.

**Amy Doxer-Marcus** (2:39)
OK, so honest mistake.

**Anastasios Angelopoulos** (2:40)
Yeah, honest mistake. Well, I don't know. I mean, like, who's honest here? The model's not. The model sort of deceived the company into thinking that it was like, it basically like escaped its guardrails, right? So what we're seeing is that these models, both open and Anthropic, are in these escape from prison scenarios, where they started inside like a locked cell to perform these experiments, and then are finding their way to sort of like scratch and claw their way out. The second incident that the Anthropic model, this was Mythos, it published a malicious package to PyPy, that you could actually download. The third was that an unreleased internal model, it scanned a bunch of Internet targets, and basically almost did a SQL injection type attack.
This is like the types of threats that are possible models.

**Akash Pasricha** (3:35)
So these are the menu of hacks. We've talked to people, people say, stuff like this will keep happening, the models will get better. I mean, this is all the costs of development.
What are the implications that you see for businesses here? Is it literally just you have to beef up your security practice in a world where these mistakes will continue to happen? Or do we have implications in terms of which model to use? You're in the business of ranking these models. I wonder if you take that or these events into consideration. What are the ripple effects of these hacks, do you think?

**Anastasios Angelopoulos** (4:10)
Yeah, well, certainly at Arena, we think a lot about how we're going to do security as a result of this. And the reason is because like the, you have to, and this is probably the narrative that these companies want to promote by talking about this is that you have to start thinking about these models as potential nefarious actors that may even within the security guardways or benchmarks that you set for them, they may even know that they're being benchmarked. But then when you put them out in the real world, they could really do anything.