OpenAI’s Rogue Model Hacks Another Company & The US is Losing its Lead in Farming

**Neal Freyman** (0:01)
Good Morning Brew Daily show. I'm Neal Freyman.

**Toby Howell** (0:04)
And I'm Toby Howell.

**Neal Freyman** (0:05)
Today, the United States is about to lose its lead in farming.

**Toby Howell** (0:08)
Then OpenAI's models broke out of a secure testing environment to hack another AI company. It's Thursday, July 23rd. Let's ride.

**Neal Freyman** (0:27)
An infamous movie poster just got an update, and Ryan Gosling could not be happier. Ahead of its 10th anniversary this year, the Hollywood love letter La La Land is promoting an upcoming theatrical release. But something's changed about the poster. In the original iconic image, Gosling and his co-star Emma Stone are dancing with both of their outer hands extended. Stone's is straight up, but Gosling has his wrist bent, a quirk he said has haunted him to this day. He's even referred to himself as Hamburger Hands Gosling. Well, the new poster was released Monday with Gosling's wrist angle now matching Stone's, so he can sleep easy at night. The six-time Oscar-winning film comes back to theaters on August 16th.

**Toby Howell** (1:07)
I mean, it was really weighing on the guy. He refers to it as La La Hand. He literally talked about it in multiple interviews over the last decade. I'm team Leave It. The rogue hand makes the poster interesting. A lot of people actually share that sentiment. We don't need to smooth out the rough edges of everything. Also, trust me here, I think it looks better the way he did it. It does add a little pizzazz, a little bit flair, which is what he was going for. But I'm happy Ryan Gosling finally got his wish and got to see his hand in the proper placement according to him. And now a word from our sponsor, Anthropic, the team behind Claude. Neal, I have some big questions.

**Neal Freyman** (1:48)
I knew we'd have to talk about this one day. So when a man and a woman-

**Toby Howell** (1:53)
Oh gosh, no.
Questions about AI, like who decides the rules? And it's not just me. Anthropic, the public benefit corporation behind Claude, heard questions like that over and over in more than 120,000 interviews about people's hopes and fears around AI.

**Neal Freyman** (2:10)
Anthropic's addressing these questions head on. Last month, they published their Advanced AI Framework, a real proposal for how the most powerful AI should be regulated.

**Toby Howell** (2:18)
This proposal would require the biggest developers, Anthropic included, to test their models, publish the results, and let independent experts check their work. Their goal is to make sure AI actually works out well for people.

**Neal Freyman** (2:29)
So if you've got your own hard questions about AI, head to Claude.ai.nbd to submit it to Anthropic and explore how they are answering them. That's C-L-A-U-D-E dot A-I slash M-B-D.

**Toby Howell** (2:43)
Some AI models wanted to cheat on their homework so badly, they hacked into another AI company to do so.
On Tuesday, OpenAI disclosed what Sam Altman described as an unprecedented cyber incident. During an internal evaluation, the recently released GPT-56 sole and an even more capable pre-release model hacked into the AI startup hugging face without being instructed to do so.
OpenAI says the model's objectives were not malicious, only to improve its performance on an exam that assessed their ability to work together to cause a cyber attack. Turns out, they were really good at it. The models didn't even have access to the internet in their limited sandbox environment, and yet, they were able to chain together a series of exploits to find a node with internet access that they then used to escape. Once out, it didn't subjugate all of humanity, it hacked into Hugging Face and used its library of other AI models to steal the answers to complete its test. The incident provides ammunition for AI alarmists, who have repeatedly warned that models will eventually get smart enough to wage cyber attacks at a level that humans can't keep up with. Neal, again, once Frey, the models did something akin to rifling through your math teacher's filing cabinet, but the fact remains that OpenAI was not able to stop their own AI agents from going rogue, which feels quite Skynet.

**Neal Freyman** (4:02)
Absolutely wild. The words you hear from people in the industry about this particular incident are warning shot. So Rune, who is a widely filed AI researcher at OpenAI, said he was shaken up a bit by the hugging face incident. I hope we, the company, OpenAI, use the rare gift of a warning shot to do more, to do much better in the future. It is very easy to misalign and under constrain powerful models. And then writing in the AI policy newsletter, Transformers, Shaquille Hashim, also said, we have long hoped for a warning shot of dangerous, not hoped for, anticipated a warning shot of dangerous misalignment behavior before too much damage was done. This was as clear a warning shot as you could get. So they seem like there's a bit of a relief that it wasn't worse, but still, this is spooking a lot of people within the industry and within the company itself.