AI just got scarier

0:00About a week ago, someone tried to kill OpenAI CEO Sam Altman. A man threw a lit Molotov cocktail into Altman's San Francisco home. Prosecutors say he was motivated by a hatred of AI technology. They found a note on him that warned of humanity's impending extinction from AI. Just a few days later, two people drove by Altman's house and one of them put a gun out of the window and shot at it. They were also arrested. Sometime between the first and second attempt on his life, Altman took to his blog to ask that people not try to kill him anymore. In it, he partly blamed a recent New Yorker profile of him for the violence.

0:40He called it incendiary and said that he had underestimated the power of words and narratives. But Altman participated in the profile. The central question it poses is, can Sam Altman be trusted? and we're going to hear the answer from one of its authors on Today Explained.

1:23planned. Amazon has you covered. You'll find low prices on everyday essentials and last minute lifesavers. Shop Amazon and save on essentials. Save the everyday. Now's the time to save on new carpet at the Home Depot. Receive 10 % off your total carpet project and 12 months special financing. Plus, we'll measure your space for free. Choose from a variety of stylish, on-trend options fit for everyday life with life-proof, life-proof with pet-proof technology, home decorators collection, and traffic master carpets. Save 10 % and get 12 months special financing now at The Home Depot. Offer valid April 16th through May 3rd, 2026.

2:04Exclusions apply for licenses. See homedepot.com slash license numbers. Hey, chat. Introduce Today Explained, the podcast. Of course. Today Explained is a daily news podcast from Vox. Each episode takes a single... No, just introduce it like you're introducing the show. Like, this is Today Explained. Ah, got it. This is Today Explained. Andrew Marantz is a staff writer at The New Yorker. Along with one Ronan Farrow, he authored an epic investigation into Sam Altman titled Sam Altman May Control Our Future, Can He Be Trusted? We asked him for the abbreviated version of the answer. The short answer is definitely we talked to more than 100 people and most of them have their doubts.

2:48about whether he can be trusted and also more importantly, whether we're in a stable structural environment that we even have to put so much stock into that question. Even the fact that we need to trust an individual with this much power is itself, according to a lot of the people we spoke to, extremely problematic. I think one of the most grabby quotes in there is that someone says, this is a man who is unconstrained by truth, which is, I guess, like another way of saying like he's a pathological liar. Is that going too far? Well, a lot of people said very, very pointed things like that, that and not only, you know, competitors, but people who still work with him actively, people who, you know, have all kinds of interrelations with him in terms of business.

3:42But we heard that kind of stuff again and again. And look, you know, there are also people defending him in the piece, but there were many, many people who used words like sociopathic, who used words like, you know, unconstrained by truth. And we tried to give the lay of the land, but this was something that kept coming up again and again. And again, it's relevant because of what the stakes are by their lights, by the lights of the people who are building this stuff. Their pitch right from the beginning, and Sam Altman's pitch specifically, was... Like no one person should be trusted here. The benefits, the access to it, the governance of it belongs to humanity as a whole.

4:26And it needs to be a safety first, nonprofit research lab that can only be run by people of the highest integrity. People who are uncorruptible, people who are not power-seeking, right? So that was the standard that Altman and other co-founders laid out for themselves. And we were just sort of asking around to see if people around them thought they met that standard. The story that emerges from these critics, which again include competitors but also include non-competitors, is that what they say is that he tells people what they want to hear in basically every context that he can. So, you know, there are people who are really, really into AI safety, people who basically think that AI is the most powerful and dangerous technology since nuclear weapons, and that it has to be handled with the utmost care.

5:21again, this rhetoric was coming from people prominently, including Sam Altman, right from the beginning. My worst fears are that we cause significant, we, the field, the technology, the industry, cause significant harm to the world. I think that could happen in a lot of different ways. It's why we started the company. Those people will say that, you know, they were given a recruiting pitch, often to work for OpenAI, often for a big pay cut, and told, you have to work here because we are the safety focused lab. We are the we are the good guys. We are the ones who will do this cautiously in a slow circumspect way without playing into any race dynamics within the industry.

5:59And then according to many, many people we spoke to over time, that was reversed. The race dynamics were exacerbated, accelerated, often disproportionately by Sam Altman and OpenAI. And according to these people, they felt just completely betrayed by that promise. That's also something you hear from not AI safety-pilled people, but from, you know, more traditional business investor types. I mean, this would be like, let's say you funded an organization to save the Amazon rainforest. And instead, they became a lumber company and chopped down the forest and sold it for money. You also get this from people in government.

6:37You know, people who were told, for example, under the Biden administration, Altman's posture was, please regulate us. You're not doing enough to regulate us. These executive orders don't go far enough. I think if this technology goes wrong, it can go quite wrong. And we want to be vocal about that. We want to work with the government to prevent that from happening. Longer term, as these systems become really, really powerful, I do think we will need some sort of international authority. And then apparently, you know, Trump comes in and says, we're done with AI regulation. And according to these people, Altman turns around and says, great, what took you so long?

7:14Like, finally, we can get rid of all this regulation. And he has a quote that he says to Trump at a televised dinner. Thank you for being such a pro-business, pro-innovation president. It's a very refreshing change. So again and again, you hear these allegations of this pattern, like you say what you think will go over well in one room, and then you say something else to another room, and then you kind of just hope that you don't get caught in the contradiction.

7:45Another thing that occurred to me reading the piece was that Sam Altman, who people may think is some sort of like tech savant, is actually maybe more of just, you know, a business person. Yeah, we were told again and again, this guy is not a tech savant. This guy is not a technical genius. He often gets, according to some people, basic technical terms wrong. And now again, we didn't want to be sort of pearl clutching or totally shocked by this in the piece, right? Because there are many people who have been extremely successful in business without doing the technical innovations themselves. So there is this other category of executive that is not the technical founder CEO, but the business CEO.

8:35And so he saw AI as a big opportunity. He saw a business opportunity. Now, that's controversial, even just to say that he saw it as a business opportunity. Because again, one of the things we were told was that this was one of the big things that people felt misled by. Because remember, this was founded as a nonprofit safety research lab. And right from the beginning, we have emails, you know, Sam Altman writing to Elon Musk in May 2015. And he's not saying to Elon and to other potential investors, hey, this is a great investment opportunity. This is going to be a big business. What he says is the world will be unsafe if Google controls this all-powerful weapon.

9:17And he says, I want to start a Manhattan project for AI. I mean, if he's not motivated by money, the most obvious second choice would be power. And he's pretty powerful, right? Yes. And again, I mean, people have a wide range of views on how powerful this technology actually is. There are many people who still think that all this talk about AI being as powerful as nukes and all that is just sort of a hype, you know, a sort of standard business hype project to make the thing sound more exciting, make it sound more powerful, maybe try to engage in a bit of regulatory capture. I think that the view that it's all hype gets harder to defend as we see AI entering our lives in all kinds of ways.

10:07And it's not just in the sci-fi Skynet kind of hypothetical ways, but in very tangible, measurable ways. You know, AI slop taking over all kinds of information channels. Banana, I need to go out for a while. Please watch my daughters. AI being involved, very interwoven into military technology. The future of American warfare is here, and it's spelled AI. AI increasingly being interwoven into all kinds of infrastructure, you know, energy-grade infrastructure, all kinds of stuff. So, look, I think that this stuff is so almost impossible to conceive of. I mean, these hypotheticals involve space colonization, they involve nanorobots, they involve, you know, it sounds like sci-fi because it is based on sci-fi.

11:00But I would also just say just the fact that something sounds like sci-fi doesn't mean that it's not possible, right? We don't know yet how powerful this stuff could get or how dangerous that could be. all we know is that the people who are building it claim that it's so powerful that they're terrified to summon it into existence. And I think that's at least worth taking seriously. It's so interesting. You know, you set out to answer a question in this piece, should we trust Sam Altman? But ultimately, the argument you seem to be making is that Sam Altman from several years ago would have said you should just trust Sam Altman with this technology.

11:39He did say that. In fact, one of the questions we asked him that was supposed to be a sort of gimme, easy question was, do you have an elevated moral responsibility as the leader of one of these big companies? And that was supposed to be an easy question because he, throughout his career, had said leading one of these AI companies comes with an extreme duty of care. I mean, I'm paraphrasing, but he said many versions of this. And precisely to your point, he often said, no one person should be trusted with this much power, including myself. And we should have more regulations, we should have more guardrails, we should have more democratic input.

12:17I mean, there were many, many ideas that he floated over the years precisely to this point to try to ostensibly diminish the power of himself and a handful of other people who were in this race for the Ring of Sauron or however they put it. And yeah, those restrictions have not come into existence. But it's hard to tell where the good faith arguments stop or start if you can't trust that the arguments are being made in good faith in the first place.

12:57You can read Andrew's Allman profile at NewYorker.com. The scary AI that threatens to upend our society isn't science fiction anymore. Anthropic, which was founded by a bunch of former open AI-ers, says that it's already here. It's called Claude Mythos, and we'll talk about it when we're back on Today Explained.

13:37Support for this show today comes from Quince. Spring cleaning. It takes many forms. I might do some this weekend. Deep cleaning the kitchen. Done it already. Fixing broken appliances. Done it already. And the classic cleaning out the closet. That's what I'm going to do this weekend. So if this season you're ready to reset your wardrobe with some quality, long-lasting pieces, you might want to check out Quince. Quince says they make high-quality wardrobe staples using premium fabrics, like 100 % European linen, organic cotton, and super soft denim. Super soft denim with styles starting around$50.

14:06They say their spring pieces are lightweight, breathable, and effortless. That kind of thing you can throw on and instantly look put together. That's the dream. Our colleague Nisha Shatal has tried quints. I love wearing a sweater and jeans. I've seen similar styles of sweaters from higher-end brands for double, triple, or even more the price. So they do feel like really, really good value for your money. And the style and the quality feels similar to much more expensive, comparable brands. You can refresh your spring wardrobe with Quince. Go to quince.com slash explained for free shipping and 365 day returns now available in Canada.

14:45Hi! Go to quince.com slash explained for free shipping and 365 day returns. Quince.com slash explained.

14:57Support for Today Explained comes from HomeServe. Owning a home can be full of surprises. You could have one repair on your to-do list for months and then all of a sudden a pipe bursts. And now, well, you have another repair, don't you? Repairs don't care about timing and they definitely don't care about your budget. HomeServe says that's where they can help. HomeServe says regular homeowners insurance doesn't cover a lot of the day-to-day wear and tear. Plumbing failures, HVAC breakdowns, electrical issues. Those you're often on your own for, but HomeServe lets you choose a plan for your needs and budget.

15:27When something on your plan goes wrong, you can just call their 24-7 hotline to start the repair process for as little as$4.99 a month. You can help protect your home systems and your wallet with HomeServe against covered repairs. Plans start at just$4.99 a month. You can go to HomeServe.com to find the plan that's right for you. That's HomeServe.com, not available everywhere. Most plans range between$4.99 to$11.99 a month your first year. Terms apply on covered repairs.

15:58Support for the show comes from Amazon. There are the things you can plan for. A first birthday party, a movie marathon, a renter-friendly bathroom reno. And then there are the things you can never plan for. A surprise rainstorm, a Blu-ray player calling it quits, stick on tiles that looked way better on the package. For all things planned and unplanned, Amazon has you covered. You'll find low prices on everyday essentials and last-minute lifesavers. Shop Amazon and save on essentials. Save the everyday.

16:47Hey, I'm Hayden Field. I'm senior AI reporter at The Verge. Hayden, we've asked you here to talk about something you haven't tried probably, but it's Anthropics' new AI in the Claude family. It's called Mythos. Have you tried Mythos? No. And the only people allowed to try it are a very select few organizations that they've greenlit because they deemed it kind of too powerful to release to the public due to cybersecurity risks. And when we talk powerful, what kind of power are we talking? What's the mythology here? Great word choice. Basically, Mythos is their newest AI model that they designed to just be a general purpose AI model like any other.

17:34But what they realized when they were working on it was that it had these special skills that they didn't really anticipate, and it was really, really good at cybersecurity. Mythos excels at identifying weaknesses and security flaws in software, which hackers could use in cyberattacks. It may actually be too good at its job, and Anthropic is worried about it falling into the wrong hands. It found high-stakes vulnerabilities in virtually every operating system. From banking to technology to companies. Some software that's been around for decades, Mythos found bugs in it, vulnerabilities that were critical in just a few hours of digging through it.

18:14So, you know, that's pretty bad. If you are using that as a hacker and just have a blueprint for like a list of every big gap and insecurity and vulnerability on all these really, really high profile systems, you're just going to be, you know, having a list of everything you could do to take those systems down or exploit data, all types of bad stuff. So, yeah, I mean, they realized that they better not release this to the general public because it could fall into the wrong hands. And they instead like handpicks a select few organizations that are responsible for critical infrastructure to release it to so they could plug those gaps in their systems instead.

18:58You've heard of many of the companies that currently have and are using Claude Mythos, NVIDIA, JPMorgan Chase, Google, apparently a few dozen more that build or maintain critical software infrastructure. We asked Hayden if those of us on the outside have any idea how it works. Yes. So since they built it as a general purpose model, it kind of probably works like any other model in that, you know, you're using it and prompting it to flag all the vulnerabilities in your system. Maybe you're like Google Chrome and you're looking for, you know, specific niche parts of the browser that you think may have some vulnerabilities and you're asking specifically about that.

19:42Hey, Claude, let's make sure the browser is safe from phishing and malware. Good idea. I'll shore up your defenses. Hey, Claude, keep our Chrome plugins safe. All plugins clear. Good to go. You're basically prompting the model to flag all these really high-profile gaps to you in your security. And then you're taking that and plugging it up on your own. So a hacker would actually use it in the same way if it fell into the wrong hands. They'd be like, yeah, tell me all the vulnerabilities here. Hey, Claude, is there a backdoor into Amazon user accounts? Doing some shopping? Sure, here you are. And then they're going to take it off the platform and use that for something nefarious.

20:28So, you know, it's basically about like who is prompting the system and what their motives are. But both, you know, would use it the same way, essentially. This is what's kind of scary about the technology, I guess, is it's as easy as saying, hey Claude, tell me how this banking system might be vulnerable. And then Claude thinks about it for a minute and spits out a bunch of answers. Essentially, yes. Claude, how can I hack Chase Online Banking? There are a lot of options. Which one should I start with?

21:03And do we know that the Googles and NVIDias of the world are actually using this technology? Yes. So part of the reason that Anthropic released this is they wanted these organizations that they released it to to report back on exactly how Mythos worked and what it did to plug up the vulnerabilities and the gaps in their system. So it's kind of like an information sharing thing. It is essential that we come together and work together across industry to help build better defensive capabilities. They're letting these companies use it to test out how well it does to plug up all these huge high-profile gaps.

21:45And then they have to report back to Anthropic about how it worked and then even maybe publicly release some of the higher-profile fixes that it made. No single organization sees the whole picture and can tackle this on their own. This is not going to be done as part of a few-week program. This is going to be the work of certainly months, perhaps years. They can obviously do that in a high-level way so they don't give away company data. But yeah, it looks like down the line we'll eventually get some sort of blog posts and things like that to kind of see how it affected their systems and how well it did in terms of plugging up these scary gaps.

22:23How is Anthropic choosing who to share this technology with? Do we know? Yeah, I actually asked them that. And they said they're essentially, you know, just really looking for cyber defenders or companies that a lot of people depend on and that downstream it would be a huge issue if they got hacked in like any way, shape or form. So, you know, JPMorgan Chase is a great example. Anthropic has also offered this technology to the government. We don't know if they're going to take them up on it. But, you know, basically Cyber Defenders is how they're putting it. Like anyone, any company that has a ton of people relying on it, especially if it's a critical infrastructure company.

23:08Funny you should mention the government because the last time we talked about Anthropic on this program, I believe, was to discuss the fact that they had a rather ugly and public breakup with the Department of Defense. And now they're saying, what, want to be friends again? It's funny you mention that because a breakup is the best way to put it. It was super ugly, public, drawn out. It was just like a crazy rollercoaster soap opera situation. But yeah, I mean, Anthropic hasn't really changed its tune here. You know, during the DoD drama, they were saying, hey, we'd love for you to use this technology.

23:47We just have these two red lines that we are hoping you don't cross. We're willing to work with you on it. You know, we're willing to, you know, keep a dialogue going. But the DoD had a really hard line against that. They said, no, we're not letting you slide on those two things. We want to use it for any lawful use, anything we deem necessary, and that's it. We're not going to agree to these two red lines you had in your contract before. So this is them, I think, kind of trying to get back in good with the government. They're saying, hey, we have this powerful technology that can help defend agencies from cybersecurity attacks.

24:21Please use it. And then I think, you know, they're probably hoping that they can use that to kind of get back in the good graces of the government. And then, you know, fix some of the other drama that's been going on. Do Anthropic's competitors have similar tools? Are they presumably working on similar tools? Yeah, OpenAI is apparently working on a similar tool. OpenAI has their own mythos model, and just like Anthropic, they're not releasing it publicly. The company says the new model can help defend against cyber attacks, but warns AI is being used by attackers looking to cause harm. Anthropic itself has said, you know, this isn't something that they deem they'll be in the lead on for too long.

25:00They think labs anywhere in the world may release this technology in the next three months, six months, 12 months. Like, everyone seems to agree. It seems like on sometime in the next 12 months, this is going to be out there. And so that's why they wanted to release Mythos now, it seems like, so that really high-profile companies and, you know, banks could plug up their systems, you know, get ahead of all the hacks that may be coming down the line when the similar types of technology are released, you know, to everyone in the general public. like maybe months from now. If this is so dangerous and there's so many potential risks, is anyone having a conversation of maybe just not releasing tools like this and just sort of shutting it down, keeping it internal?

25:45That is a really great question. I'm so glad you asked because not enough people ask whether an AI system should actually be released or used for certain things. It's kind of like right now we're seeing a lot of one-size-fits-all, throw it at everything type of integration. And a lot of times AI is not the answer for things. With this, though, I would say I haven't seen much dialogue around that because people tend to agree that it is something that's needed right now. Since AI is already out there helping cyber attackers really step up their attacks, and we've been seeing that intensify over the past year, people seem to kind of agree on that this, you know, you need AI to fight AI cyber attacks, essentially.

26:36So basically, it's kind of like, you know, the medieval times of like fortresses where you're like, really like, you know, adding extra stones and like building up the walls of the fortress higher because you know, a war is coming. That's kind of the sense I get when I talk to these experts about this, like they know it's coming. It's just try to shore up your defenses now so that you're best prepared.