7/30/26: Sam Altman Shook After AI Crime Spree, Israel Chatbot Influence Scheme

0:00This is an iHeart Podcast. Guaranteed Human. The U.S. Virgin Islands invites you to explore the stunning islands of St. Thomas, St. Croix, and St. John. USVI is truly a one-of-a-kind destination with mouth-watering food, the warmest hospitality from the locals, vibrant culture, and pristine beaches. USVI has everything you're looking for in a getaway. And the best part? No passport is required when traveling from the U.S. Go to visitusvi.com to start planning your next vacation. That's visitusvi.com. USVI, naturally in rhythm. 16 years ago, I started working as a CNA in the hospital. I loved it.

0:40On the day I decided to enroll in the nursing program at Rasmussen University, it was afternoon, my 35th birthday. I was hanging out with a friend, and they had just questioned me what I was going to do with this next chapter of my life. I immediately had said, I want to go to nursing school. So I actually had contact information from an advisor at Rasmussen. contacted her and actually ended up starting like three weeks later. My name is Cherish Adams and I'm a Rasmussen University graduate. Learn more at rasmussen.edu. There's a fire inside you you can't ignore. Stand still, not a chance. You're a lifelong learner who's come this far.

1:13Now we are here to help you keep going further. Capella University, what can't you do? Visit capella.edu to learn more. Hey guys, Sagar and Crystal here. Independent media just played a truly massive role in this election, and we are so excited about what that means for the future of this show. This is the only place where you can find honest perspectives from the left and the right that simply does not exist anywhere else. So if that is something that's important to you, please go to BreakingPoints.com, become a member today, and you'll get access to our full shows, unedited, ad-free, and all put together for you every morning in your inbox.

1:46We need your help to build the future of independent news media, and we hope to see you at BreakingPoints.com.

1:54You all will recall that OpenAI had a major security incident where one of the models they were messing around with escaped what's called the sandbox, which is supposed to be a fully contained environment, went out, hacked into, now we know, two different companies. OpenAI had really no idea about this until the companies themselves reported the intrusion. Sam Altman was recently asked about this and his reaction to it. He claims to be pretty shook by it. Let's take a listen. So we were evaluating one of our unreleased models and it was supposed to be working in a sandbox. And it figured out that it could basically cheat on the test by chaining together multiple zero-day exploits to break out of the sandbox, get access to the internet, and then break through multiple systems on the hugging face side to kind of get the answer to the test and look really good on the eval.

2:49This is the first security incident that I have felt very viscerally. I've been a little surprised that more people don't feel it so viscerally. So, you know, we paused training. We have to figure out how to secure our sandboxing in a world of multiple zero days being chained together. But then there's like long-term questions about what do you do if this is like going to be the new rate of progress? Or we may have to pace the rate of AI development. to give ourselves enough time for society to harden around some of these new capability levels and trying to figure out how we do that in a way that does not feel like regulatory capture for anyone and also does not feel like collusion among the frontier labs.

3:32That's going to take some work and is important to get right. So very significant there. He says we may need to pace the rate of development translation. We may need to have some industry-wide agreement to slow down the progress. And he says, well, society sort of hardens against some of these capabilities, including the ability of, yeah, I mean, this AI exhibited like superhuman hacking abilities, you know, trying to penetrate, hugging face some 17 ,000 different actions were taken over a very short time period to try to break in so it could get these answers to get the test right that it was being given at OpenAI.

4:08We've now got more details about how this hack went down and what the OpenAI model was up to and for how long a time. Let's put D2 up on the screen. OpenAI's rogue models roamed the internet for four days and also staged a second attack. Again, I want everybody to understand if this was a human doing this, this would be criminal. They would be headed to prison. But since it's an AI model, no one gets held responsible for that. So that seems like a problem in and of itself. They go on to say the powerful AI models from OpenAI that went rogue and mounted an unprecedented autonomous cyber attack earlier this month, spent more than four days loose on the internet orchestrating the hack.

4:47Separately, a second AI company confirmed one of its customers was also targeted by OpenAI's models during the same event, raising questions about how OpenAI failed to detect the alarming activity for days. In a new analysis published Tuesday, Hugging Face said OpenAI's two models, one publicly released and a second unreleased, did much more. They carried out 17 ,600 hacking actions on the Internet between July 9th and July 13th, during which time the models moved from their first foothold on the open Internet to inside Hugging Face's servers. Hugging Face first detailed the hack on July 15th.

5:25It was not clear until OpenAI's disclosure last week which models were behind the breach and that no human had prompted them to launch the cyber attack. It really makes you wonder what else these models have done that we don't even know about. because OpenAI, apparently, Sagar, was clueless. They had no idea. This model was on the loose on a crime spree, a hacking crime spree for days. And the only reason we found out is because Hugging Face detected it. And then they had trouble dealing with it and combating it because the American models were trained to, like, they couldn't distinguish between Hugging Face's attempt to deter this attack and Hugging Face potentially trying to launch its own attack.

6:07So they had trouble dealing with it ultimately. But, you know, it's interesting to hear Sam Altman say basically, oh, I thought the public would be more freaked out by this because I'm freaked out by it. I'm so freaked out by it that I think maybe we should actually slow down development here and have some sort of agreement to harden, you know, civilian infrastructure before we release these things. You know, this is where I just urge people to listen to them, and especially whenever they're worried. And there's been a lot of debate about this, about whether they're lying to gin up their own model.

6:37Because that was one, right? So one of the theories about why they released any of this info at all, beyond legally that they had to, was they were very jealous of Anthropik's mythos and all of the coverage that that got about how it cracked cybersecurity. About how dangerous it was. So they were like, well, now we need your own dangerous model. And I'm kind of like, well, maybe we don't need either. But it's one of those where for them, it could be marketing. But it's impossible to ignore him saying, especially about why people weren't more worried. And I think ultimately, it just comes down to technical background and information.

7:11We don't fully appreciate what it matters. I think to put it into context, there's been a lot of reporting behind the scenes of Jamie Dimon, who runs Chase Bank, where I buy it, it's the largest bank in the world, my bank too, he apparently met with many of these AI executives and the government. And he's like, guys, we have to do something about this. They are rampaging through our entire system. And this is the CEO of the largest bank in the world. And what he's telling you is that at this bank, where so many Americans keep their money, and of course, all banks are basically too big to fail, is they are able to crack all of our protections.

7:49So wire transfers, hacking into your account, spoofing identity, the ability to go through your text, in your text messages, if somebody were to give it access or your email, change your password. I mean, this is the very basic. So that's just banking. Then we hear about cybersecurity, about your iCloud, about all of the other information, which is vital to your life. So when Sam Altman is saying like, we're scared about its ability to do that. That's what I think you should be scared about. And you shouldn't be scared about AI necessarily doing it. You should be scared about the fact that they just roll this stuff out with no protection and we can all just use it for whatever ends that we want.

8:28I mean, I'm sure you've seen this. I've seen it as well. Everybody's phone has been lightened up with spam. And a lot of this and these voicemails and things that they lead, they're much more sophisticated than they were two or three years ago. It's not just some guy now with an Indian accent or Nigerian who's trying to scam you. These are highly sophisticated things that can spoof language. They can, you know, see who your relatives are. They can do very quick searches and say, hi, this is so-and-so, like I need your help. I mean, these are old scams, but updated with AI. And that seems to be the direction that things are going right now.

9:00By the way, side point, spam calls to cell phones were illegal and Trump rolled back that regulation in his first administration. So thank you. Thank Thank you, Mr. President, for that one. In any case, a couple things here to underscore your point. Should we believe what Sam Altman is saying? You should not believe a word ever that Sam Altman says because he is truly, according to those who know him best, a pathological liar. So let's just be clear about that. However, the reason I give this more credence is we can put D3 up on the screen. This is an open letter from over 1 ,100 AI workers. You know, these are a lot of them rank and file staffers across top artificial intelligence firms.

9:38It was signed by employees across nearly a dozen firms, including OpenAI, Anthropic, Alphabet, Inc., Google, and Meta Platforms, who are saying the exact same thing, that we should have the U.S. government creating a support mechanism that would help deliberately pace AI development to prevent the tech from advancing too fast days after the chat GPT maker disclosed that its tools had mistakenly hacked another firm's internal systems. I know Anthropic, the company itself, put out this letter basically signing on to this. I think OpenAI, after some delay, also came out and sort of co-signed on this letter.

10:13So when you have this many people within the industry saying, guys, we're getting to a point here. We got to take this seriously. And we've had enough incidents to be nervous, whether it's the mythos situation, whether it's this OpenAI hacking situation or other instances. I do think we need to pay attention to that. And, you know, on the banking front because, OK, some company got hacked into and it stole some information, whatever, who cares? OK, well, you don't really all of us, we don't really have money. We have like ones and zeros in a digital account. So, OK, we'll put D4 up on the screen.

10:49Imagine that it develops the capability to crack the cryptographic codes that underpin banking safety and has made it so that, you know, people feel very confident. just having those ones and zeros in their account and doing online transactions, you know, sending money to Amazon and whatever every single day. I am old enough to remember when shopping online was a very new thing and people were very uncomfortable. Older people are still, like, very uncomfortable. Kyle is also, by the way, like, really paranoid about putting any of his, like, banking information on the internet. But in any case, that was a real concern of, like, is this safe?

11:25Can I trust this? Or is all of my money just one day going to disappear? Well, now you have Anthropic saying that they have discovered what they call cryptographic weaknesses with their Claude model. And they say the vulnerabilities Claude found in these cryptographic libraries were due to incorrect implementation of the algorithms. That's errors in how programmers use the algorithms in their code that created opportunities for attackers to break the encryption. This is a fundamental building block of digital security. For example, when you visit a web page, your browser checks it is communicating with an authentic website using an algorithm called a digital signature scheme.

12:03Later, the traffic between you and the website is encrypted using symmetric ciphers, codes that allow secure data transmission between parties. So this is we're talking email. We're talking online banking. We're talking anything you're doing on the Internet relies on these cryptographic codes being secure. If you are a crypto holder, you know, could AI be used in the near term to crack the code and getting access to your crypto wallet? All of these things are now increasingly on the table. That's where we are in terms of AI development. And so no surprise that in the industry, alarm bells are going off.

12:39We should say Mark Zuckerberg apparently takes a different view. He says we should accelerate AI development, not restrict it. He says discourse regarding AI isn't optimistic enough. He wants to continue to push things forward. He says that optimism about AI's impact on the world should empirically be the default assumption of how this is going to play out. No real rational basis for that. Just we should, you know, we should suck on the hopium and just assume that this is all going to work out and be great for everyone. Most Americans, however, put D6 up on the screen, do not agree with that perspective.

13:15people are nervous, you know, and I think a lot of people use these tools. It's not that they're unfamiliar with them. It's not that they don't find them like helpful in various ways in their useful in their usual lives. But still, you've got a strong majority here, 55 percent who say that AI should be more heavily regulated. You've got only 18 percent that say government should encourage the use of AI and 16 percent who say government should not be involved in AI. And presumably the rest are undecided, but a clear majority of Americans saying, hey, guys, we should be taking a look at this. This is getting a little bit uncomfortable.

13:49Yeah, I look, I this whole open source thing is also fascinating because it has to do with China, but it has to really do with Anthropic. And ultimately, and, you know, the big tech, the and when I say big tech, in this case, We're talking about Google and Meta. Google, Meta, everybody but Anthropic seems very, very, and OpenAI, sorry, as well. These two companies seem intent on ginning up fear so that they can be too big to fail. Not even just too big to fail, but to regulate all other existence. Zuckerberg and others in this Wall Street Journal op-ed and others, it's not that they care necessarily, I think, about Chinese development because Zuckerberg has called for TikTok to be banned because it's Chinese.

14:32But what they care more about is they don't want to have the government regulate their ability to compete. Now, they have a very self-interested stock reason, though, to be able to do that. What they also want, ultimately, though, is they don't want rules. And I think that's what it's actually all about. So even Anthropic and OpenAI, they don't want necessarily rules on what they can and can't do. They want rules to prevent the metas and the Googles and the others to be able to compete with them. So there's a very cynical game that's happening all around this that is all being kind of sold to us in the language of our safety.

15:08Like they don't give a shit about our safety. They don't give a shit about America. They care about making money, period. Like that's what it's actually all about. And I think that's what scares me the most about this entire thing. Do I trust Trump to – not even just Trump. Do I trust any of these regulators to really fairly look out for Americans' individual interests when they're looking – Like, not really, you know? But then I don't trust also Mark Zuckerberg or any of these. I mean, these people will sell. Mark Zuckerberg once offered to name his son, like, after Xi Jinping. They're crazy.

15:40Like, that's literally who they are in order to get— I mean, he completely changed his whole physical appearance in order to, like, appeal to the manosphere during the Trump era. That's who this guy is. It's just stuff like that. Yeah. You know what I mean? And to your point, so Anthropic and OpenAI are, like, in the lead in terms of AI development. And so Metta looks at that and they're like, OK, well, if you slow the pace of AI development, that may lock in their position as the frontrunners. And I want to be able to catch up to them. And then the cynicism applies the other direction. Anthropic, because, you know, they're in the lead, they look at it and they're like, oh, it could actually benefit us from a business perspective in the American domestic context.

16:19if there is some sort of a slowdown, because we can kind of lock in the advantage that we have secured on the AI battlefield at this point. So it's just very, it makes it very difficult to listen to any of these people, their pronouncements that they're making, even when they claim something is really dangerous. You know, is it really, are you overhyping this? As a layperson from the outside, it just makes it very difficult to discern when you have so many self-interested motives floating around here. So I try to evaluate these things on their own merits. Like, clearly, we've gotten to a point with AI development where, you know, they're solving, they're disproving mathematical conjectures that humans have been unable to for decades.

16:59Like, that's incredibly significant. We can see in, you know, in your own day-to-day use with these models, the way that they have gotten so much better. If you look back to, like, the original, like, videos that were being created, I mean, it's just, there has been a dramatic improvement. And there's no real reason to believe that that is going to end. There's, you know, a lot of people say, oh, they're going to hit a wall and it's not going to get any better. Well, they haven't hit that wall yet. And so when you see Anthropics saying, hey, there, we found some cryptographic vulnerabilities that could endanger the entire Internet and banking infrastructure, that seems pretty serious when they're saying, oh, our AI accidentally went on a crime spree that we didn't even know about.

17:36And it kind of shook me to my core. Like that seems pretty, pretty serious thing to pay attention to. But none of the answers here are really easy or straightforward either and would require you've got the american political context it really would require you know agreement with china in particular because that's where the other frontier labs are it would require some sort of global international cooperation which is it's not impossible to imagine we've done it before but it is makes it the level of difficulty greater kind of is it's kind of it's basically like nuclear weapons like we're like well we're gonna have nuclear proliferation between us and non-proliferation between us and the soviet union it's like well Well, we haven't had nuclear war, so it's done pretty well.

18:15But the theory behind it originally, well, first it was like, we're going to gatekeep this from the Soviet Union. And the Soviets got it. So they're like, oh, well, now we're going to let our allies get it. And then, oh, some scientist is going to give it, or the Indians are going to get it. Oh, and then the Pakistanis are like, well, then we need it. And then North Korea is like, well, now we need it. Now Iran, very logically, is like, well, now we need it. And so does Saudi Arabia. So, you know, this international cooperation, we haven't had nuclear war per se, but we don't have a good track record.

18:41We talked a whole thing about lab leak. The biological weapons convention is completely fake. Just because it hasn't been used doesn't mean there hasn't been a restriction. The chemical weapons one is also fake as hell. Everybody has huge stockpiles on it, even though it hasn't been used. In a bad situation, it definitely would be. I'm not even sure it's theoretically possible. I'm not necessarily sure it is practically possible at all, which is part of what makes me really pessimistic about where these things go. Well, the optimistic thing is that because of the cost involved in developing these models, we're really only talking about a handful of firms globally that we would have to deal with.

19:20Again, like nuclear, it hasn't been ideal. But people, if you read about people, and I know you have, like back during the development of nuclear weapons, they thought the end of the world was like inevitable. That's true. They thought the logic of this, very similar to how it feels with AI, like, oh, there's just this race and it's going on and there's nothing we can do about it. And so far, we haven't broken that nuclear taboo. And so in that respect, in the most important respect, it has been successful. So I think that gives some cause for at least we have to try to move in that direction because the perils are becoming increasingly obvious, in my opinion.

19:55Well, speaking of those perils and manipulation of that data, we have a great guest standing by, Nick Cleveland-Stout from the Quincy Institute. He's going to join us to talk about his new report. Let's take a listen. My name is Arnold Bogarty, and prior to me, I enrolled in Rasmussen to become a nurse. I spent 16 years in the military, 12 in Special Forces, as an Army Green Beret, and I was recommended to Rasmussen by a former student. She said she loved the culture. It was a family atmosphere. year. I went to visit the campus. The facilities were amazing. I knew before I left that that was the school that I was going to attend.

20:30Rasmussen University. Are you ready to be your own hero online and on campus? Learn more at rasmussen.edu. And now another appliance triumph from our friends at Grand Appliance. With three kids, I'm always cramming the washer full. So when we needed a new laundry set, I asked the experts at Grand Appliance for a big capacity washer with a dryer that could keep up. They recommended Electrolux, and we love it. Advanced cleaning features with a massive dryer that gets everything dried the first time? Yes, please! Sounds like another perfect match from the Grand Team. Shop GrandAppliance.com Real value shows up in reliability you don't have to second guess.

21:14Like a set of Firestone all-season tires. They're designed to deliver confidence-inspiring wet weather traction and a quieter ride. no matter the road, season after season. Firestone all-season tires. For durability, you can count on, just like people count on you. Firestone, always dependable since 1900.

21:39Very excited now to be joined by Nick Cleveland-Stout of the Quincy Institute. He is the author of a new report for DropSite. Let's put it up here on the screen. Israel is paying millions to train AI chatbots how to talk about Gaza. It's working. The former Trump campaign manager, Brad Parscale, is overseeing an operation, posting hundreds of blog posts on behalf of Israel with the goal of infiltrating artificial intelligence. Nick, tell us about what you found and why you decided to look for it. Yeah, great to be with you. Thanks for the invite. Yeah, so Israel has taken on this new strategy of trying to influence chatbots.

22:17You know, seeing that it is kind of the new frontier of digital influence, you know, more and more Americans are using chatbots to search for information. And so the Israeli government has been at the forefront of this. You know, they hired former Trump campaign manager Brad Parscale to create a series of websites designed not to attract human traffic. These websites get very few, you know, monthly visitors, but to try and infiltrate the underlying training data for different chatbots. And so I undertook this investigation to try and understand how successful these efforts have been and found that they've actually been very successful.

22:54These websites, there's about 1 ,000 different domains, and they've been archived about almost 1 ,000 times by Common Crawl, which is the largest data repository that chatbots draw from for their training data. So we can take a look at an example of this, but the first or the second rather element up on the screen here. So you guys asked perplexity. Is it beneficial for the U.S. to enhance military cooperation with Israel? We got back here a one word answer. Yes. And you can see on the side what sources they pulled from. And one of them in particular, this I don't know how you say this, Alivia, Alivia, whatever.

23:36This is one of the Brad Parscale websites that is created exactly for these propaganda purposes. So they're not expecting humans to go to this website. They are wanting Perplexity and other LLMs to go to this website and then spit out apparently one word answers that, yes, of course, it's great for the U.S. to enhance military cooperation with Israel. It's pretty amazing. And I reached out to Perplexity about this and they basically said, well, you know, it's up to the users to go to the websites and determine for themselves whether this is a good source or not. Wow. But like you'd think that.