Anthropic recently took its two newest AI models offline due to an export control order from the Trump administration, prompting broad debates about AI policy and digital sovereignty.
On the latest episode of TechCrunch’s Equity podcast, Sean O’Kane, Rebecca Bellan, and I discussed what actually prompted the administration’s moves against Anthropic, and what this might mean for the broader AI ecosystem.
As Sean put it, “Anthropic has not had the best relationship with the Trump administration in a way that stands apart from the other leading AI labs,” so perhaps other Anthropic’s rivals don’t need to worry about a similar crackdown.
But Rebecca also noted that leading cybersecurity experts have “signed an open letter to ask Trump to revoke the order, and they say it’s actually dangerous to have to pull these advanced cybersecurity capabilities from network defenders in the U.S.”
And we wondered whether this could all end up being good publicity for Anthropic, especially since — in Rebecca’s words — “everybody loves a bad boy.”
Keep reading for a preview of our conversation, edited for length and clarity.
Rebecca Bellan: As I’m sure many of our listeners know, the U.S. government basically just forced Anthropic to pull its two newest models offline — Fable 5, and then there was also Mythos 5, which was the one that was available to current Mythos users, [whereas] Fable 5 was more available to the public.
They sent a letter [last] Friday that cited “national security concerns.” No one knows what those concerns are. That report has not been made public, they gave no specifics and told [Anthropic] that they had to ensure that those models couldn’t be used by any foreign nationals. So Anthropic was like, “Okay, I guess we have to just pull the models entirely, because we don’t know when someone’s a foreign national. A lot of our own employees are foreigners.”
But really, [reports said] the White House got tipped off to this because of some Amazon researchers that allegedly found a way to bypass Fable 5’s guardrails. Amazon CEO Andy Jassy raised these concerns with the White House, and it just kind of spiraled from there.
Sean O’Kane: This all moved really fast, especially for a Friday afternoon into a weekend. And it’s at the same time that the administration was ostensibly trying to negotiate some sort of treaty for the war that it started in Iran.
Rebecca: Friday evening for us in New York. They love a distraction.
Sean: Let’s step real far back for a moment. Anthropic has not had the best relationship with the Trump administration in a way that stands apart from the other leading AI labs — I think there’s an element, at least, of that playing here.
So do you think that this is going to have implications for those other companies? Do you think that the Trump administration would be less inclined to sort of turn off the tap on one of those competitors?
Anthony Ha: Part of the context here is that both the reporting and an analysis from independent security experts suggest that the actual security risk from Anthropic is not that unique. So a lot of this seems to stem as much from parts of the Trump administration and Anthropic just [not getting] along very well. Whatever risks there are, those things are gonna blow up out of proportion just because it seems like they can’t have a civil phone call with each other.
If you’re another company — on the one hand, maybe that’s advantageous to you, because you can say, “Well, we just don’t get these guys mad at us and we can do what we want.” But that’s also not a great regulatory landscape to just [say], “Boy, I hope they don’t get mad at us.”
Rebecca: On the one hand, it definitely feels retaliatory — after the government labeled Anthropic a supply chain risk, there’s this big lawsuit going on between them, it really feels like the White House is just looking out for any excuse to pummel Anthropic. And I feel that way not only because that was my initial reaction, but because of what a lot of cybersecurity researchers have said. They say that this should never have triggered an export control [order]. They’ve all signed an open letter to ask Trump to revoke the order, and they say it’s actually dangerous to have to pull these advanced cybersecurity capabilities from network defenders in the U.S. Anthropic itself said some of the same jailbreaks could have been found in several other AI models.
Cynically, it’s like: Okay, are you just pausing Anthropic so that others can catch up to where Anthropic was?
But at the same time, I’ve also seen reactions that [say]: Anthropic kinda had this coming. They’re like, “This is too dangerous for anyone to use, but not us, we’re the good guys.” They’re talking out of both sides of their mouth. A week before Fable came out, they were [saying], “Hey, we need to slow down AI, guys. It’s getting really dangerous.” But then boom, “Here’s our most insane ever, super powerful model, go off.”
Anthony: In some ways this feels like a microcosm of a lot of the discussion around AI, where people like Sam Altman and Jensen Huang are [saying], “Hey, let’s try to lower the temperature. Why is everybody mad at us?” Well, you spent the last couple years essentially saying you’ve built this God machine that will take jobs away from everyone. It’s not exactly a shock that people don’t feel great about this.
And there’s something about the way Anthropic talks about Mythos in particular, where they’re like, “This is the most incredibly powerful model ever, it’s too dangerous to release to the public.” And so on some level, [you say,] “Well, okay, let’s say that we take that seriously then. That means that there’s going to be an incredible level of scrutiny around it.”
And I do wonder — it does seem like Anthropic is not happy about this. I want to be careful about not overstating how this could be beneficial to them. But we also ran some stories about Ramp analysis to highlight the fact that the last big blow-up between Anthropic and the Trump administration was good for the company, in at least some ways. Downloads of Claude shot up. I think a lot of people who maybe had thought of ChatGPT as the chatbot, the AI assistant before, suddenly they were looking at Claude as maybe the more responsible one, the more “resistance” one.
And in the same way, [while] Anthropic is very stressed out about this, this could, again, make their models seem even more powerful.
Rebecca: Definitely. “We’re so dangerous.” Everyone loves a bad boy, right? Everyone’s like, “It’s the most powerful model, even Trump says so. Of course, I’ve got to get my hands on it.”
To pull this off, a very basic HTML page was turned into a series of UTF-8 encoded bytes that were then declared to be a standard PNG image. The original 208 byte payload plus 4-byte PNG header only used part of a 9×9 pixel favicon. With a larger favicon image as typically used you could thus easily store more data, whether as visual noise like here or a bit more hidden.
You must be logged in to post a comment Login