Why Anthropic Shut Down Its Best AI (Special News Update)

0:00Whether you're traveling or just at home, your online activity shouldn't be up for grabs. Proton VPN adds a layer of protection that keeps your browsing habits private and your data out of the wrong hands. I use it to change my virtual location so I can stream my favorite TV shows while traveling abroad for work. I don't have to worry anymore about falling behind and seeing spoilers while I'm away from home. Proton VPN keeps what you do online private and lets you access the internet like it should work. Open, secure, and on your terms. Unlike most VPNs, Proton is backed by strong European privacy laws and years of expertise creating a safer, faster, and more open internet for everyone.

0:38So whether you want to watch content from anywhere, get around blocked sites, or just have added peace of mind while using public Wi-Fi, ProtonVPN has you covered. It's easy to get started. Right now, ProtonVPN is offering our listeners 70 % off a two-year plan when you go to ProtonVPN.com slash uncanny. That's P-R-O-T-O-N-V-P-N dot com slash uncanny for 70 % off your two-year plan. That's ProtonVPN dot com slash uncanny. NBR believes in your right to be curious about the big questions and the unexpected ones. That's what Shortwave, their science podcast, is all about. When we stop to ask questions, the world gets more interesting.

1:22Shortwave spends each episode answering one big one, like, why do we have nightmares? How does AI affect my energy bill? And how can trash be good for the environment? Science isn't just for the PhDs of the world. It's for everyone, including those of us who didn't realize we love science. In under 15 minutes, Shortwave follows a question wherever it leads and brings you along for the ride. It's science, but not the kind you remember from school. It's surprising, playful, and full of wonder. News podcasts are great, but sometimes you want to experience the world in a different way. shortwave gets you out of your head and in touch with the world around you personally i loved hearing about janet jones horse trainer and neuroscientist teach me about horse and human relationships throughout history follow npr shortwave podcast and exercise your right to wonder when you need to build up your team to handle the growing chaos at work use indeed sponsor jobs it gives your job post the boost it needs to be seen and helps reach people with the right skills certifications and more Spend less time searching and more time actually interviewing candidates who check all your boxes.

2:33Listeners of this show will get a$75 sponsored job credit at Indeed.com slash podcast. That's Indeed.com slash podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed sponsored jobs. Welcome to Wired's Uncanny Valley. I'm Zoe Schiffer, Director of Business and Industry. I'm Brian Barrett, Executive Editor. Today on the show, we're bringing you a special, shorter episode to explain the latest escalation between Anthropic and the Trump administration. Over the past weekend, the Trump administration unexpectedly imposed export controls on Anthropic's latest models, Fable 5 and Mythos 5, basically ordering the company to bar access to, quote, any foreign national, whether inside or outside the United States.

3:15In response, Anthropic disabled access to these models for all users to make sure they could comply with the directive. Now, anthropic executives are reportedly in Washington, D.C. to speak with the Trump administration officials and figure out how they can resolve this impasse. So this is remarkable. It's also it's our first emergency podcast. So very exciting. This news unfolded on Friday pretty late. What was your reaction when it came down? It came past your way. I mean, it was unprecedented. I don't think we've seen export controls applied in this way. Obviously, this is just the latest escalation in an ongoing fight between Anthropic and this administration.

3:58But it also seems like it comes on the heels of Anthropic really drumming up a lot of hype around how powerful and scary its models are. And so there was a little bit of irony of like, wow, you tell everyone that your models are really, they can do incredibly capable things. They're too powerful for their own good. And eventually people will believe you. This part of the story goes back to April or so. So Anthropic has this model called Mythos, and they say, look, it's so good at cybersecurity things specifically, like finding bugs in code and creating exploits for it, that we cannot release it to the general public.

4:40We are going to give it to a few trusted partners so that they have time to defend themselves before we release it to the public. Very scary stuff. And look, people have said different, there are varying views on how essential that was. Is that fair to say? Yeah. So I think actually we need to back up even further because the myth is released. I know. It came after Anthropic had gone head to head with the Department of Defense, basically saying that it wanted guardrails on how the military could use its models. The Department of Defense did not want those guardrails in place. And so there was a standoff.

5:18And by the time they released Mythos, it kind of looked like, well, whether this was the intention or not, it was getting Anthropic back in the administration's good graces. And we talked to sources on a variety of different levels who said that while relations between the Pentagon and Anthropic were still pretty icy, relationships between Anthropic and the intelligence community were actually starting to thaw. So it looked like these talks had been successful and at least the White House was in communication with Anthropic as a result of the Mythos release. So then we get to the point where Anthropic says, okay, people have had time to steady, ready their defenses in light of this powerful model we're going to release.

6:03They go ahead and say, all right, we're releasing this on June 9 as Claude Fable 5. They say, hey, we're putting this in public, but we are making it so that you cannot use it for cybersecurity. You cannot use it for biological weapons. weapons. I think there were a few restrictions they put on it. And they said, look, we tested this up and down. We red teamed it, which is the phrase for you treat it as though you were an attacker and you can't jailbreak this thing. It's solid. It's not going to do the bad things. Go forth. Exactly. So basically they were releasing a mythos class model, but with a bunch of safeguards in place.

6:41And what that meant is that if you asked the model questions about how to create a bioweapon or what have you, the model would basically refuse to answer. Yep. Can't do that. But then after the release, Amazon CEO Andy Jassy calls the administration and basically flags a jailbreak that would essentially allow someone to bypass the model's guardrails and potentially get it to do bad things. Jassy wasn't the only person to make this call, but he's the only one we know about right now. Yeah. And obviously, we don't know exactly how that unfolded, but it's Interesting that it came from a competitor.

7:17A competitor who is invested in Anthropic. So it's a complicated web. What happens next is pretty widely disputed between the White House and administration officials and Anthropic. But basically, the White House says that they call Anthropic and try and get CEO Dario Amadei on the phone. But they were unable to reach him because he was supposedly at a wellness retreat. Anthropic vehemently denies that this was the case. Either way, he'll need one after this. Yeah, for sure, for sure. Regardless, they get on the phone, looks like, you know, an hour and 15 minutes after the request was made to have that call.

8:00And Dario basically tries to get more information about the jailbreak, like what was going on, how serious was it, why is the White House flipping out? It looks like Anthropic didn't believe this was as big a deal as the White House believed it was. And there was some disagreement about the severity of what they had found. In any case, the administration asks Anthropic to voluntarily remove the models. And Dario, importantly, doesn't commit to doing that. It doesn't look like he said no, but they get off the call and there's not the commitment. And so then the next turn is that the administration imposes export controls.

8:37says you can't let anyone access this who is a foreign national but of course there's no real way to guarantee that unless you just shut them down altogether so zoe that's our timeline and there's a lot to unpack here the thing that i'm sort of immediate my mind immediately went to this and i still kind of struggle with is the underlying logic of what is going on here so So if mythos class frontier models are too powerful to be released to the general public, unless you can promise that they aren't actually going to be jail, you can't jailbreak them. Where do you go from here? Because eventually all these frontier models are going to be at mythos class or better.

9:18They're all going to be this powerful. And there's really one of Anthropik's arguments was, look, you can't make a model that's entirely 100 % jailbreak proof. Like you can prevent sort of big universal jailbreaks. But if you're looking for like little things that are going to break here and there, you're not gonna be able to fix that. So where does the U.S. go from here? I think that it's still very much an open question and one that every frontier AI company is asking itself right now. We know that OpenAI and the other labs, they're having serious discussions internally about like, well, what do we do?

9:55Even if they haven't gotten, you know, a similar directive from the Trump administration, like all of these companies are releasing new models all the time. And I think there's a lot of fear in the Valley right now about like, you know, how do we ensure that we don't end up in the same situation as Anthropic? I will say a couple of things. One, I think obviously given the history that we already discussed, Anthropic has a much worse relationship with the Trump administration than basically any other company. And this is definitely a line that Anthropic sources are hinting at. Like, this wouldn't have happened to someone else, basically.

10:27This is because, I'm paraphrasing completely, but, like, this is because they hate us. Yeah. And then I think there's this question about, like, well, was the jailbreak a—like, was it really, really serious? Or was it something that was a little more tailored and narrow and the administration didn't have, like, the technical expertise to, like, understand? I think, you know, the administration is saying, look, Amazon brought this to us. We ran it by like cybersecurity experts up and down the administration, and they said it was really serious. So like, we believe we have, I think the word was proof, but Anthropic is still seemingly disputing that.

11:05Because what Anthropic is saying is like, look, the stuff that you can, the capabilities you can unlock with this jailbreak are already available on other models that have been out there for a while. Like you're not introducing any unique risk with this specific jailbreak that you don't already have in some of these other models. And that sort of ties into some of the most vocal defenders of Anthropic were cybersecurity professionals. There was a big open letter from cybersecurity folks and tech CEOs saying, look, you know, however worried you are about this, we're more worried. Because while this, while Mythos class models are good at offensive hacking, they're also really good at defense.

11:50And we need that help with defense because there are other models out there that are going to help people attack us. It's an interesting sort of tradeoff that I don't know that the administration has really fully weighed. No, and I think that this is, you know, some of the pushback we are seeing is like, how are we supposed to compete with China when China is releasing really advanced open source models? Anyone can use them. And in the U.S., we can no longer use, generally speaking, like the most capable models we have access to. And I don't think the administration has a really strong answer to that, at least not yet.

12:26This is also coming at a time when AI policy within the White House is being pretty hotly contested, even within the White House. Like David Sachs, who was the AI and crypto czar, is no longer in his kind of official position. Susie Wiles and Michael Kratzios are kind of like running part of that show. But I think there's a lot of disagreement, even within the administration, about the right way forward. Are expert controls the right move? If not, how do we control these things? How much should the administration be weighing in on models before they are released? Because this is happening shortly after Trump released an executive order saying that they wanted AI companies to give them a preview about the—they wanted AI companies to voluntarily give them kind of a preview about the most capable models before a general release.

13:19At the same time, you've got Dean Ball, who's briefly in the Trump administration, who is an AI policy expert, never shy with his opinions, said it was, quote, simply cartoonish that an administration that supports exporting advanced AI tips to China would at the same time say no U.S. allies, no foreign allies could have access to its best models. Expert controls are generally applied to like hardware. This is kind of a novel use of them to apply them to a software tool. That also means that the outcome of this is going to be sort of uncertain, right? There's no real precedent here all around.

13:56What do you think happens from here, Zoe? Like, what are the next steps? We know anthropic officials are trying to meet with the White House, going to hammer it out. But what's the short term and I think more importantly, the long term implications? I would be surprised if they don't reach some sort of agreement that results in Mythos and Fable being generally available. Maybe they'll have more guardrails in place. Maybe there will be some sort of formal review process. But I would expect that there is just a lot of demand for frontier AI. And the idea that in the United States, people cannot access models that are as powerful as what people have access to in China, I don't think that's going to be a palatable solution.

14:35I also think that every frontier AI company is scrambling right now to figure out how they can make sure they have good lines with the Trump administration, that the Trump administration is aware of what models they have coming and that they maintain the appearance of being on better terms. Because Anthropic has really been winning the kind of PR war on a lot of fronts by saying we're the safe AI company. We were willing to go head to head with the Trump administration in order to like keep people safe and not let them, you know, have unfettered access to our AI for military purposes. We won't put ads in our model.

15:16Well, now there's an opportunity to basically make them look like they're being messy, that they're moving quick, that they don't have good ties with the administration. And I think that that's an opportunity that the competitors are not going to let pass them by. What does it mean for, we talked about open source in China. There's some open source happening in the U.S. and efforts there, not nearly as robust and further behind. But does this have implications in terms of, you know, at least if you're an open source model, you're sort of not as subject to these conditions or terms or anything.

15:54You're just sort of out there. Does this give open source a boost here to know that you are not going to be subject to the whims of the administration necessarily because, you know, no one has control over it in that sense because it's open source? I don't know. I mean, it does seem like the United States has kind of ceded open source to China. And, you know, this has been something you and I have talked about extensively. I think if you look at the latest models from China, it seems pretty clear that they have been able to distill, extract the weights from U.S. models and then create really capable open source models that have very, very similar capabilities.

16:34I actually think that's a huge vulnerability for U.S. AI companies. If they're able to continue releasing really impressive, capable models, perhaps it won't matter. But if that, you know, rate of change slows and then you have DeepSeek offering, for example, a very, very capable model that looks a lot like Claude and it's cheap, if not free, then, you know, I think that that could be a huge problem. But I don't know. It doesn't. It seems like there's just so much competition and we've really leaned into closed source at this point that I think it would be genuinely hard to see that change. Zoe, so much of this conversation, both literally the one we're having right now and also just broadly is about things happening at the CEO level and the heights of the administration and the strategic level.

17:25What does all of it mean for your average Claude user, someone who uses Anthropic? How does this impact them? What are the implications of them going forward in terms of their AI usage? My feeling is that it doesn't have a huge implication for the average user. I think the reality is that most people are using Cloud and ChatGPT like Google. They're just using a fraction of the capabilities of these models, which makes sense. These models, at the end of the day, they're like open text boxes, and they require a lot from a user in terms of what do you want to do and knowing how to kind of direct them effectively.

18:09So I think if you're, you know, the average person and you've been using Claude, I honestly don't think you'll see a huge difference. You probably weren't on like the bleeding, cutting edge of like pushing these models and what they can do. And you will continue to like have a good experience. I think the like implication years down the road in terms of like what this means for US AI and like the trickle down effect for users, that still is an open question. But what do you think, Brian? No, I agree. I think unless you were planning on making a biological weapon, you're probably not going to feel too much of a pinch, especially the Mythos class model had only been out for a few days before it got yanked.

18:49So even if you had started using it and found, oh my gosh, this can do things that I didn't know AI could do for me, which is not really going to be the case for almost anybody, then yeah I agree I think you're not even going to notice or even be vaguely aware of it but yeah long term it's a little bit of a mess and that mess is going to have implications for how AI gets developed in the United States when things get released how well it keeps up so you know I think not much for now and long term I think it's a bigger question of like do people do does the average person need AI that's much more powerful in any case and how long will it take to get there.

19:29Yeah. Ultimately, I think like really powerful AI is important for the people building the layer of apps, like on top of the models, you know, it's like maybe you could do really cool things that would then turn into interesting consumer products and those will require like better and better models. But I think right now we're still seeing people kind of just access the chatbots and use them, you know, to cheat on their homework or whatever. I do think there's just, I mean, I don't want to get us too far off topic, but there's an interesting split happening right now because we're hearing more and more about how unpopular AI is.

20:03You and I have talked about like the college students booing commencement speakers who bring up the technology on stage. And then there's the fact that usage of AI has like absolutely skyrocketed in recent years and doesn't appear to be tapering off. Like people hate the technology overall. They use it all the time. And they also use it in a kind of surface level way that might not be pushing the absolute bleeding edge of what the models are capable of. Yeah, I do. I will take a little bit of a sideways approach here too, though, in that if there is an impact on the average cloud user, or just the average internet user, I do think there is something to the fact that, look, hackers have access to all these open source models and they are using them.

20:52So So, you know, does that mean it's more likely that things get hacked in a world without Mythos being available to cybersecurity defenders? Maybe. I think it's a little bit overheated. I think that there's a lot of defense in place already, but maybe. I mean, maybe there's some weight to that. When people like Alex Damos, who was then Facebook's head cybersecurity guy for years and years, comes in and says, hey, this is actually a problem, I'm inclined to listen.

21:27That's our show for today. We'll link to all the stories we spoke about in the show notes. Uncanny Valley is produced by Kaleidoscope Content. Adriana Tapia produced this episode. It was fact-checked by Daniel Roman. Pram Bandy is our New York studio engineer. Mark Leda is our San Francisco studio engineer. Kimberly Chua is our senior digital production manager. Kate Osborne is our executive producer. and Katie Drummond is Wired's Global Editorial Director.

22:01Uncovered windows can make your home feel up to 20 degrees hotter. Stay cool and save up to 50 % off custom window treatments during the 4th of July mega sale at blinds.com. From outdoor shades to room darkening blinds, finding the perfect fit is easy. Get free samples, expert design help, and professional measure and install services, or DIY with confidence and support every step of the way. Shop up to 50 % off site-wide, plus huge savings on doorbusters, right now during the 4th of July mega sale at blinds.com.