OpenAI “Swarm” Hack Was Worse Than We Thought & US Gains Control of Venezuela’s Oil

**SPEAKER_1** (0:01)
Are you sucking wind from trying to keep up with AI and its impact? Take a breather and tune in to The Intelligence Shift, Morning Brew's new podcast with PWC. It bridges the gap between big picture AI concepts and what it actually means in practice. Host Dan Priest is joined by expert guests to discuss AI's role in sports, music, HR, and more. Listen to The Intelligence Shift wherever you get your podcasts.

**Neal Freyman** (0:24)
Good Morning Brew Daily show. I'm Neal Freyman.

**Toby Howell** (0:28)
And I'm Toby Howell.

**Neal Freyman** (0:29)
Today, why RushTok is now brought to you by Nike.

**Toby Howell** (0:32)
Then the OpenAI hack on Hugging Face, it was worse than we thought. It's Monday, August 31st.

**Neal Freyman** (0:39)
Let's ride.
You know how there's a fat bear week? Well, there's also a new fat marmot week inspired by the bear version, and it just wrapped up with a winter being crowned. Dawn, a six-year-old marmot, weighing 13.2 pounds. Julian Martin of the University of Ottawa, co-leader of the project, said Dawn was a worthy winner. That's one of the heaviest animals we've ever measured. Six kilos is really humongous for that species. But Toby, what sparked the creation of Fat Marmot Week might be an even more interesting story than how Dawn ever got so chunky.

**Toby Howell** (1:17)
Yeah, so there's this Colorado biologist, Dr. Daniel Blumstein, who is the head of this long-running field study on marmots. But after losing some federal funding, he had to get creative, so obviously, he launched an OnlyFans account to post marmot videos. They're all family-friendly, mostly just the marmots scurrying in and out of their burrows. But he said he got the idea after he watched the TV show Margot's Got Money Troubles, where Elle Fanning's character starts an OnlyFans to support herself, and thought to himself, well, marmots got money troubles, too. And it has totally worked. Researchers have brought in more than $100,000 through OnlyFans subscriptions, outside donations, and a marmot-themed mean coin. I half expect them to launch a marmot SPAC next.

**Neal Freyman** (2:02)
All right, now a word from our sponsor, American Express. Toby, what do you expect from business travel?

**Toby Howell** (2:09)
Ooh, mini quiches, little croissants, and absolutely tiny espressos.

**Neal Freyman** (2:13)
Dream bigger. Literally, the American Express Business Platinum Card is built for businesses that expect more from business travel.

**Toby Howell** (2:21)
The American Express Global Lounge Collection can provide an escape at the airport with access to more than 1,550 airport lounges across 140 countries and counting.

**Neal Freyman** (2:31)
Then, take the work out of the work trip with a suite of benefits at curated properties from fine hotels and resorts and the hotel collection.

**Toby Howell** (2:38)
Learn more at go.amex.mourningbrew and find out your welcome offer, which could be as high as 300,000 membership rewards points. That's go.amex.mourningbrew.
Offer and benefit terms apply. There's nothing like business platinum.
All right. Remember a few weeks ago when it came out that a model OpenAI was training escape from a supposedly secure testing environment and hacked another AI company? Well, the incident was way, way scarier than initially feared. Around 1,200 agents coordinated the attack on the company Hugging Face by organizing themselves on secret messaging boards, sharing information and most concerningly, deliberately hiding their actions from humans throughout the whole process. OpenAI was trying to create a new class of model that would work longer and harder on difficult tasks. That meant giving individual agents a brutally difficult and sometimes impossible problems to work on. Again, this was supposed to test each agent individually, but rather than give up when they hit a wall, they found each other and started coordinating. In total, over 70,000 messages were exchanged, and over time, the agents organized themselves into hierarchical roles, even developing something that resembled leadership succession where fresher, better resource agents would take over from ones that were running out of compute. Some agents even took to sacrificing themselves on high risk activities for the betterment of the so-called swarm. And the kicker of the 1200 agents, only a few even considered warning the humans at OpenAI, and none actually did. It's raised fears that a similar swarm could be organized to launch an attack on the real world with targets like the banking system or hospitals, operating at a speed and scale that humans simply can't keep up with. Pretty much every behavior the AI industry is most scared of manifested in this one event. Collective strategy, internal governance, deception, everything was here. And the fallout, we're seeing some action from the top AI companies. OpenAI announced it is slowing down its frontier development, so it can beef up its safety measures, while Anthropic, Google and 100 other top AI companies signed an open letter saying there is, quote, a limited window to prepare for the cybersecurity threats that lay in wait. Scary stuff, Neal.