Claude Mythos 5 を巡る争い：Anthropic がトランプ政権と輸出規制で対立

As the rest of the country celebrated the USA’s first World Cup win and the New York Knicks championship, Anthropic spent its weekend fighting the Trump administration over its latest model release. At 5:21 PM on Friday, the company received a US export control directive to suspend access to its Mythos 5 and Fable 5 AI models by “any foreign national” inside or outside the US, “including foreign national Anthropic employees.” The only way that was possible, Anthropic determined, was to completely disable products it spent the past week hyping — and travel to Washington, DC in hopes of changing President Donald Trump’s mind. Now, over the coming days, the US government could dramatically alter the trajectory of the entire industry, dealing a major blow to American AI companies.

Claude Mythos 5 and Fable 5 are built on the same foundation as Anthropic’s Mythos Preview, which Anthropic dubbed too dangerous to publicly release. (The company’s warnings could be seen as genuine concern or more hype for their own model — or both.) Mythos 5 was made available to a select group of government agencies and companies, while Fable 5, which featured additional safeguards, was deemed “safe for general use.” But when a report indicated those guardrails may have failed, Anthropic’s dire warnings about Mythos falling into the wrong hands came back to haunt it.

A source familiar with the situation, who participated in the negotiations between Anthropic and the Trump administration, said the administration called the AI lab on Friday around 1pm ET and gave the company a 90-minute ultimatum to shut down access to Mythos 5 and Fable 5. If it didn’t, then the government would impose export controls on Anthropic by authority of the US Commerce Department.

The source said that Anthropic executives were talking to the White House within 15 minutes of that first call, confirming that CEO Dario Amodei joined the discussions about an hour and 15 minutes after that initial call. Amodei directly spoke with US Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, and National Cyber Director Sean Cairncross, in some cases more than once, the source confirmed.

Anthropic wrote in a release on Friday that the company believed that the government “believes it has become aware of a method of bypassing, or ‘jailbreaking’ Fable 5.” Rather than an existential threat, though, Anthropic said that the jailbreak in question was a “potential narrow, non-universal” one that was “shared with the government” by an entity the company declined to name. Moreover, Anthropic said the behavior wasn’t unique to Fable 5. “We have reviewed a report that we believe is the basis of the government’s directive and validated that the level of capability displayed there is widely available from other models (including OpenAI’s GPT-5.5),” Anthropic wrote.

*Semafor* reported, citing one source familiar, that the hubbub began because the US government was concerned that a China-linked group had accessed the technology. But the source said that the China rumors went back weeks, referring to a large global telecommunications company that was initially cleared to be included in access to Mythos Preview, and that when the US government shared its concerns, Anthropic immediately revoked access.

An X post by David Sacks, the US government’s former AI and crypto czar who stepped down in March, didn’t mention China either. Sacks did, however, mention the unnamed entity that had exposed the issue to the government, calling it “a highly credible trusted partner of both Anthropic and the USG who was testing Fable [which] came forward with a jailbreak of those guardrails.”

Some reports point to Amazon CEO Andy Jassy as the person who flagged concerns to the US government after researchers at Amazon had red-teamed Fable 5. That conclusion stands at odds with some independent red-teamers, who have said they were impressed with the level of the protections.

The source familiar with the negotiations said that the Amazon research was explicitly mentioned in conversations with the US government. The person added that Anthropic had had access to that paper within days of the Friday export control directive and had been going back-and-forth since then with Amazon researchers to discuss it.

Everything in that paper, the source said, could be achieved by OpenAI’s GPT-5.5.

Anthropic spent the weekend scrambling to make nice with the Trump administration, beginning with virtual meetings and then flying employees to DC, including Dave Orr, Anthropic’s head of safeguards; Logan Graham, who runs its frontier red team and has led work on Project Glasswing; and Nicholas Carlini, a leading frontier developer and cybersecurity researcher. *Axios* reported, citing a source familiar with the Trump administration’s thinking, that the company simply has repeatedly made missteps in its communication with the administration and that it “has not done a great job at trying to speak to the administration and appreciate the ideological differences.” For Anthropic, the timing couldn’t be worse: the company had banked on Mythos to help it recover, in part, from months of high-profile clashes with the US Department of Defense.

The source familiar with the negotiations said that Anthropic pre-briefed the administration on Fable 5, and that the US Department of Commerce conducted testing pre-deployment, with no concerns shared at the time. The source added that Anthropic had been working closely with government agencies since Mythos Preview’s release.

The Trump administration initially took a hands-off approach to AI safety — but post-Mythos, it has become more ambivalent, even as it frets over the threat of losing the AI race to China. Now, prominent cybersecurity leaders have warned that sidelining Mythos 5 and Fable 5 could give China a significant AI advantage. Trump’s move has galvanized international calls for alternatives to American AI systems, while effectively putting a major US AI company’s new flagship model on ice.

A public letter from tech and cybersecurity executives called for restrictions on Fable 5 to be repealed on Sunday. “Not all of us agree that AI regulation is the right way forward,” the letter states, adding that if regulations are going to happen regardless, then they should be rooted in “scientific evaluations developed with input from industry and academia.”

Alex Stamos, chief product officer at Corridor, told *The Verge* he organized the public letter because the countless number of vulnerabilities in the past decade-plus, written in a variety of different coding languages, require AI to patch before bad actors find them. “We’re in a race, and I think policymakers don’t understand that,” Stamos said. “There’s this weird arrogance, this idea that American labs are hugely ahead of our adversaries that will always be true, that it’s really important to restrict access because of that. I just think that’s foolish. If the labs are ahead, it’s only by a matter of months. And you can see that in the open evaluations. The cutting-edge models are only something like six months ahead of the Chinese models — and those are the models we know about.”

The public letter goes on to state that though Anthropic’s Mythos-class models are skilled at finding cybersecurity vulnerabilities and taking advantage of exploits, they aren’t “uniquely good” at these tasks and that Fable 5’s safeguards “were so aggressive as to be the source of humor in the cyber community on launch day.” Stamos told *The Verge* that “there’s a real overstatement of Mythos’ capabilities. Anthropic is somewhat responsible for this themselves, clearly … Mythos is great, but the real turning point was really last year.”

Stamos said the industry is awash with backup contracts being signed with non-US companies and open-weight models being deployed on alternative hardware arrangements because the past weekend made political risk part of companies’ business plans more than ever before.

“They are laughing at us in Beijing right now,” Stamos said. “One of America’s champions is being kneecapped by the US government while we’re in a race with the Chinese. It’s just incredibly stupid. That’s why I wrote the letter, and I think that’s why a lot of people signed onto it.”

Ben Van Roo, co-founder and CEO of Legion Intelligence, a system of agents for the national security community, told *The Verge* that “the directive of ‘no foreign national should use this model’ is the most impossible thing to enforce.” He added, “When I first read that, my whole… [network of] AI community nerds was exploding.”

To make matters even more urgent, OpenAI, Google, and Microsoft have all come out with their own comparable products to Anthropic’s Mythos, making many of the same claims about their effectiveness and risks. If the Trump administration bans Anthropic’s advanced cybersecurity models, it can make a case for banning its competitors’ models, too. That could spur AI industry leaders to unite and help out Anthropic or, as with its fight over autonomous weapons with the Pentagon, position themselves as a safer and more compliant alternative.

Even as the Trump administration is trying to free tech companies of regulatory hassles, the Anthropic order could amount to a dramatic restriction on powerful AI models — depending on how the next few days play out.

Legion Intelligence’s Van Roo called it “uncharted territory” in the regulatory setting, adding that he doesn’t think this is the last time something like this will happen.

We’ve also entered the era of AI populism, when a growing number of people are pushing back against the AI industry’s outsized influence and the concentration of power at the top via data center protests, pledges to quit using AI chatbots, lawsuits over wrongful deaths, and even attempted attacks on AI company CEOs. Van Roo says the Trump administration’s recent moves against Anthropic could stoke “greater fears and concerns, potentially for the wrong reasons.”

The source familiar with the negotiations described the weekend’s conversations as constructive, with some members of the administration admitting that putting export controls on model providers isn’t ideal, since competitors with similar products may find themselves under the same restrictions — and since the US government is currently exploring a program that would encourage the export of American AI systems.

Monday’s talks concluded with no resolution as of yet.

As Anthropic continues to negotiate with the US government, there’s little chance that the company’s other myriad issues with the Pentagon won’t come up — namely, the ongoing battle between Anthropic and the Department of Defense over acceptable usage policies for Anthropic’s tech by the US military.

“This is new and we’ve never had anything potentially this drastic before, and it does have some real ramifications” in terms of how to enforce access to powerful models, Van Roo said. “Who gets to use this new technology that continues to outpace our own ability to regulate it?”

2 Comments

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

• Hayden Field

-

-

-

-