It’s an AI company valued at almost a trillion dollars, but Donald Trump is clipping Anthropic’s wings. The US president effectively banned the company's state of the art Claude AI models Mythos 5 and Fable 5, citing security concerns. But how founded are those concerns and is it problem that the US government can control the release of groundbreaking AI?  Today, Sharon Goldman, journalist and founder of Ground Level AI on what happens next. Featured:  Sharon Goldman, journalist and founder of Ground Level AI Subscribe to ABC News Daily on the ABC listen app.Program:More from ABC News DailyAustralia, United States, AI, Cyber Security, Information Technology Industry, Technology, Business and Industry RegulationTranscriptSam Hawley: It's an AI company valued at almost a trillion dollars. Donald Trump is clipping Anthropics' wings. The US President effectively banned the company's state-of-the-art AI models, Mythos 5 and Fable 5, citing security concerns. So how significant is it that the US government can simply cut off access to groundbreaking AI? Today, Sharon Goldman, journalist and founder of Ground Level AI, on whether the ban is really justified. I'm Sam Hawley on Gadigal Land in Sydney. This is ABC News Daily.Sam Hawley: Sharon, the Trump administration has had a fractured relationship with Anthropic. There's a sort of love-hate thing going on, isn't there?Sharon Goldman: There is, absolutely. That's a good way of putting it.Sam Hawley: And at one point, Donald Trump actually described Anthropic as a radical, left, woke company. And earlier this year, he ordered any business working with the military to stop using it. So just tell me what happened then.Sharon Goldman: Okay, so this was back in March. The government had already been working with Anthropic. The Pentagon had been using its family of models called Claude for a variety of purposes. And the Pentagon wanted certain restrictions removed from Anthropic's existing contract, including restrictions for domestic mass surveillance and autonomous weapons. And for Anthropic to agree to a new contract, Anthropic refused. This was back in March. And in response, the Trump administration cancelled the deals, designated Anthropic a supply chain risk. And that's really quite a label. It's really historically been reserved for companies tied to foreign adversaries. But as you said, it means that no defence contractors can use Anthropic's technology in the completion of any work for the Department of Defense.Sam Hawley: Yeah, and a huge blow that was for Anthropic.Sharon Goldman: Right. They went to DC and tried to negotiate. They sued. But in April, Anthropic then introduced Claude Mythos, caused a stir because it was so skilled at finding and exploiting software vulnerabilities. People said it was too dangerous to release to the public. Mythos was initially restricted to select organisations. But then at the end of May, Anthropic announced that they had a new model coming out that was a Mythos level model. And on June 9th, they released Claude Fable 5. And this was like a model at the same level of Mythos, but they said that it didn't have the same cybersecurity risks. But just a few days later, the U.S. government shut it down.News report: Anthropic says it will abruptly disable its most advanced AI models for all users...News report: The suspension came after the U.S. government took the unprecedented step of prohibiting the use of the models by any non-U.S. national.Sharon Goldman: They hit Anthropic with, you know, what was really unprecedented export controls, saying that they had to cut off foreign access to this Fable 5. Apparently, it was due to a warning from Amazon CEO Andy Jassy about cybersecurity fears, about the idea that there had been a jailbreak when Amazon was testing it. And so, the net effect was that suddenly the U.S. government had cut off access to this other model.Sam Hawley: Okay, all right. So hang on, just so I've got this straight. Anthropic, it brought out Fable 5, which basically had more safeguards in place than its Mythos model.Sharon Goldman: That's right.Sam Hawley: Which it had to kind of shelve because of safety concerns. And hundreds of millions of people, according to Anthropic, were actually using Fable 5. Tell me, what does Fable 5 do exactly? I never actually used it.Sharon Goldman: I never used it. Well, that's not true. I did. It was there for a few days. I mean, it was just, it was part of the family of models that you could choose from. And so, whenever the next level of model is coming out, you know, every developer who's coding with Claude Code or wanting to use the most sophisticated new model is on it. So, they are testing it. They're using it for their projects. So, there were tons of people already using Fable, and then suddenly it was gone.Sam Hawley: Yeah, and Anthropic was given just 90 minutes notice by the Trump administration before this new order came into force.Sharon Goldman: That's right. And they immediately posted a blog post saying, you know, like, this is unprecedented. And they really pushed back saying that this was not something that should be shut down. If you were going to shut us down, you should shut everyone down then because this model is really no different in that sense than any other model.Sam Hawley: And Anthropic says that really the only way it could comply with this order was to disable the latest models for all customers. It wasn't just for people overseas.Sharon Goldman: Exactly. Like, how would you really tell who's who? I mean, it just was impossible. So, they just took it down.Sam Hawley: So, Sharon, Anthropic has had these models basically banned because of what the industry calls jailbreak. Now, just explain jailbreak. It basically is a user tricking the model into bypassing safeguards, isn't it?Sharon Goldman: That's right. So, an AI jailbreak in this case would be used to trick an AI to kind of bypass its built-in safety guardrails. And you can sort of manipulate the prompt in various ways. People have tested all sorts of things. You can bypass these sort of alignment rules that would normally prevent the model from generating harmful content or instructions for malware, say, and compel the AI to produce outputs when it was supposed to refuse those outputs. But jailbreaking, I've been told by many experts, is not something you can fully ever get away from. You have to manage it.Sam Hawley: Right. Okay. But the Trump administration is, what, concerned that it could lead to cyber attacks.Sharon Goldman: Right.Sam Hawley: Although Anthropic doesn't think that is the case, right?Sharon Goldman: That's right. That's right. So, that's where you kind of wonder, like, you know, what's the real problem here? Yeah. But supposedly, from what I've read, the negotiations have not led to the Trump administration allowing the model to go back up again.Sam Hawley: I mean, there's mixed messages, right? Because in an interview this week with the Axios show, Donald Trump was really changing his tune on whether Anthropic is a threat or not.Marc Caputo, Axios host: Do you view Anthropic, and to a degree its CEO, Dario Amodei, as a threat to national security?Donald Trump, US President: Well, not now, but a week ago, maybe. He responded to us very quickly. Because, you know, it's tremendous liability. People get put in prison immediately for that. You know, you can't play games with that. And he responded very responsibly.Sam Hawley: It's not really clear what's going on, is it?Sharon Goldman: It's not. And I think that's why there's just a lot of uncertainty out there. We're seeing that other companies like OpenAI are kind of waiting before they put out their next models, because no one's really sure what's happening. Again, it's a little hard to tell whether this really is an issue of jailbreaking that, you know, deserves to be shut down, or what exactly Anthropic could do to satisfy the government and re-release the model.Sam Hawley: Yeah, and before this ban was actually introduced, the Anthropic boss, Dario Amadei, he was on Bloomberg saying that he really does want to cooperate with the government.Dario Amodei, Anthropic CEO: Hopefully, six months or a year from now, we have a much more secure internet ecosystem than we had in the past. We're trying to get to that world. And we're doing the best we can to open up mythos to new cyber defenders. We've been talking to the government. We're very respectful of their recommendations. They're slowing the pace at which we open it up because they're worried about counterintelligence risk. I think that's sensible.Sam Hawley: But remind me, Sharon, what is the worst case scenario when it comes to these really advanced AI models? How can they be misused?Sharon Goldman: I think the thing that people in the government are really concerned about are attackers getting into critical infrastructure, for example. You know, whether that be our electrical grids, you know, our critical water infrastructure, and an attacker could start a military operation against the US. Certainly, there are other folks in AI that are really worried about more existential risks that attackers could come up with biological weapons and attack a country. You know, all of these military and government concerns are real, for sure.Sam Hawley: All right. Well, Sharon, when the US stops the export of AI models, of course, nations, including Australia, miss out on this advanced AI technology. It is a sign, isn't it, just how powerful America is when it comes to this technology, that it can just cut other nations off whenever it wants to.Sharon Goldman: Yeah, and that's been the other part of the story that I think has been so fascinating, of course, where other countries outside of the US are continuing to wake up to the idea that, wow, we need our own sovereign AI, or we need to access, you know, open models, or our own domestic models to have less dependence on, you know, like one US vendor, because it certainly seems like this could be some kind of weapon in a way, a security weapon, or this kind of ban could become sort of a regulation and use access to influence who gets frontier AI, which is concerning, I can imagine, to countries outside the US.Sam Hawley: Yeah, because building our own AI models, I mean, it's incredibly expensive. Many nations wouldn't have the expertise to do that. I mean, China and the US are leading the AI race. It's pretty hard for other nations to get started right at this moment, I would have thought.Sharon Goldman: Well, that's what's so interesting, because China is known for their AI prowess, particularly in the open source world, which means that their models are not closed down, like you don't need to have a subscription to use them. Anyone can use these models and sort of build upon them. And China has gone this route, some say, to kind of enhance their own soft power. And if you're offering your own open models to the world, then more of the world will get on board with Chinese AI, for example. Instead, the US has more so gone towards this closed model idea, which Anthropic and OpenAI and mostly Google are following.Sam Hawley: And banning models into the future. I mean, the US could do that for any number of reasons, right? And it doesn't just have to be security concerns. It would give America a competitive advantage militarily, but also in every aspect of society, from health to finance.Sharon Goldman: I mean, I do feel like, you know, there are many other models that can be used for all sorts of things. But if you're looking for the most sophisticated frontier models, then yes, what you say is right. That for certain very important military use cases, for example, where you would want the best of the best, then the US could sort of hold that over other countries.Sam Hawley: So what do you think, then, Sharon? Will the Trump administration keep these bans in place? Or in the end, really, does it need to just let these companies do their thing?Sharon Goldman: I think what people are hoping is that there will be a way for Anthropic to show that they have, you know, satisfied the government, you know, not by eliminating all jailbreaks, but, you know, having stricter access, government approved testing, maybe identity controls, you know, all these cybersecurity layers that I was saying, like this defence in depth. And maybe that would become a best practice that other models would follow. And there would just be more of a process for these things to be rolled out. But I think people are saying that it's still going to take a little time. I feel like there's some bets going on of how long into July this is going to go on, for example.Sam Hawley: Yeah, okay. But in the end, these models, they are becoming incredibly powerful, aren't they? Should we be concerned about the dangers that might come with that? And should governments, including the Trump administration, really be looking to regulating this further?Sharon Goldman: Yes, I mean, I definitely think most experts agree there needs to be some kind of regulation. The trick has always been figuring out what that should be. I think what's happening right now in cybersecurity in particular, really shows how fast this is moving, and how governments can no longer wait to, you know, noodle over this too much longer, because the more powerful models keep coming. So when Fable 5 is allowed to come back, then I think the government will have to, you know, sort of be forced to put into place more of these best practices, more of these regulations, really.Sam Hawley: Sharon Goldman is a journalist and founder of the independent publication Ground Level AI. This episode was produced by Bridget Fitzgerald. Audio production by Sam Dunn. Our supervising producer is David Coady. I'm Sam Hawley. Thanks for listening.