The new AI model that’s alarming Washington | The Economist

Mythos is a new AI model trained by. Anthropic. The reason why it's causing a. fuss is that Anthropic say it is an. extraordinarily. competent cybersecurity engineer. It can hack things really, really well. Um. to give you an example, Opus 4.6, the. the previous top-tier model from. Anthropic, the one that the public can. use, is quite good at finding weaknesses. in technology. It found a vulnerability. in a version of Firefox that has since. been fixed. But if you ask it to then.

use that vulnerability to hack a. computer, it falls down. Anthropic tried. hundreds of times, they got twice a. working exploit on the very first step. Mythos, 181 successful exploits of that, and then 27 went further and actually. built a working attack chain that. affected the Windows registry, a real. like deep-level attack. That's a step. change, right? It's not just sort of a. couple of percent higher on a benchmark. And it raises the risk that anyone using. Mythos becomes a top-tier hacker, even.

if they don't have any tech capability. themselves. Okay, Zanny, you've been in. New York and Washington. How much is. this Mythos moment causing alarm there? A lot. I mean, it's interesting. I got. here basically a week ago, so just after. Mythos was announced by Anthropic, and. I've been talking to a bunch of. government officials, I've been talking. to tech leaders, I've been talking to. business leaders in bigger gatherings. and one-on-one. And I've kind of watched. over the past week as the alarm has. grown. There is real alarm in the.

hitherto very hands-off Trump. administration. Secretary of the. Treasury Scott Bessent and the head of. the Fed Jay Powell summoned in the banks. for an emergency meeting last week, and. I've really sensed as even in in other. parts of the administration now, people. are going, "Oh my god, this is a real. wake-up moment." As people are realizing. the potential risks involved in a model. like this. And it's kind of it's been. sort of alarming and interesting to. watch as pretty much everyone I've. spoken to as the days go on has. mentioned Mythos within about the first. 5 minutes of our conversation. This is,

you know, being perceived as dangerous. by Anthropic itself. What's been um. Dario's response to that? So, they have. released it behind closed doors. This is. actually not the first time Dario has. done this. In 2019, OpenAI had made. GPT-2, right? This large language model. And we were all very excited cuz it. could produce good text. And they. decided not to release it for 6 months. after they trained it because they were. afraid of what it might do on the. internet. That decision was probably. wrong in hindsight. Like, it upticks in. spam and fake reviews weren't perhaps.

worth uh holding it away from the. public. This time, I think there's. really something there. They they have. kept it available only to themselves and. 1,100 picked partners. Massive companies. like Apple and Microsoft and the Linux. Foundation, which makes the open-source. operating system. And the idea is the. hope is that these companies will be. able to use Mythos to fix their products. before any capability like Mythos makes. it to the public at large. Now, Zanny, you've met Dario recently and he.

he he doesn't mind the PR. So, I mean, and this has got him a a lot of. publicity. Is is this for real in your. view? And is it just about cybersecurity. or might it go broader than that? So, I've met Dario a number of times over. the years and he has long been of these. uh you know, AI gods the most publicly. focused on safety. And he has always. said uh you know, this is going to be. very dangerous. We need to have. government regulation. We need to be. very careful. And as Alex said, you. know, he held back an earlier model.

He's worried about bioweapons. He's been. absolutely And so, there are a lot of. people, particularly his competitors, who are saying, "Oh god, crying wolf. again. This is just marketing. And you. know, the other models are quite close. behind and they're just making a big. splash about this cuz it's good for. them.". But I my sense is that, you know, this. time really is different. And I have. talked to members of senior people in. companies that actually now have mythos. and they are all, you know, corroborating the idea that this is.

actually extremely powerful. So, I do. think Dario Amodei is definitely one. who's focused on safety. Of course, this. is this is good for him and so I'm sure. there's not a little little no little. bit of self-servingness about it, but I. think it's actually serious and for. real. I think that leads us to the next. question, which is, okay, there's a. danger here, bad things can happen, there is the beginnings of a of a. response to it. I want to look now at. how how systematic these sort of. protections can be. And Danny, I I want.

to start with you. I remember David. Sacks, who when he was in administration. was, um, you know, a booster for a kind. of laissez-faire approach. Let them. cook, let them carry on, he used he used. to say. Do you think the administration is ready. to intervene? Well, I think I think the. short answer is yes, because I think. they have really been freaked out by the. power of mythos. But you're right, it's. a very big shift from an administration. that came in basically pooh-poohing the. Biden administration, which was very. focused on trying to create kind of a.

regulatory framework. And in comes the. the Trump administration, absolutely. not, they're accelerationist, they want. to go as fast as possible, unfettered. competition. And so, the question is now. how do you do it? And I think that it's. not just what they want to do, it's the. how little time there is. Mythos exists. And so, my sense is that we are going to. have kind of informal actions very fast, which will involve continuing this. approach that the most powerful models. are first released only to a small set.

of, you know, trusted companies. That. kind of limited release, I think will be. rolled out because it also suits the. companies. The government is going to. get involved and I think we'll basically. say we need to see these models, we need. to know what they're and we're going to. have a say on how how things are. commercialized. And then I think we can. So the question is what happens. thereafter? And probably, and this is. what people are talking about, we will. it will evolve into some kind of. industry-led. sort of certification. approach, which will be the kind of big.

model builders getting together with the. government saying, "Okay, this kind of. uh model is all right for release." The. kind of tradeoffs in this are huge. Have. the heavy hand of government, then. America falls behind, you don't get the. benefits of this extraordinary. technology. That's bad. Go too slow and. you have an AI disaster, an AI accident. And as you know, Ed, I have always. thought that the kind of race dynamic. between America and China and between. these companies meant that we wouldn't. get any change in the Trump. administration's approach until there.

had been an AI accident. And I think. Mythos may actually be the wake-up call. before an AI accident, but it was pretty. close. So, Sonny, that that leaves me. feeling pretty alarmed. I mean, in other. words, we've got, you know, a small. space. in which to begin to get talking and. take this this this issue seriously, begin to get governments involved. But. it is a it is a very short time. You. should be alarmed. I mean, I you know, you you absolutely should be alarmed. This is an extraordinarily alarming. moment. However, I'm going to, you know,

with my perennial optimism, offer you one reason to perhaps be be a. bit hopeful. Firstly, I do think this. has been a wake-up call for this. administration, which had been so. extremely hands-off. The question is. whether they are kind of competent. enough to work out how to deal with it. with with Anthropic and others. And. remember, just a few weeks ago, they. were, you know, furiously having a row. with Anthropic and and, you know, deeming Anthropic this company a supply. chain risk that couldn't work for the. Pentagon. So, you know, it it's going to.

demand some cooperation. But the other. is there is this meeting coming up, the. summit uh between President Xi and. President Trump in May. And I will wager. um I'm completely speculating here, but. I will wager that this subject will be. discussed. Because I actually, even. though, you know, America is focused on. being ahead in AI, I think there is a. recognition that there are some things. that it is in no country's interest to. have. It is in no country's interests to. have the capabilities of taking down. critical infrastructure in the hands of.

some crackpot somewhere. And so, it I. think we will get the beginnings of some. conversation about how you can have. coordinational standards, because that. is essential for any approach to be. lasting. I think it'll be done in an environment. of massive mistrust. I'm not putting a. huge amount of weight on it going. anywhere, but I do think when you have a. moment like this, people start thinking. differently.