THE ECONOMIST: Australia locked out of Anthropic’s most advanced AI models as Trump tightens security

“It is time to go beyond transparency to more serious and binding regulation of AI,” wrote Dario Amodei, the boss of Anthropic, a leading AI lab, on June 10. His firm had just released Fable, a constrained version of its world-leading Mythos large language model, which could perform superhuman feats of hacking, among other dangerous tasks. What followed on June 12 is not what he meant. For the second time in four months, the Trump administration has dropped a bombshell on Anthropic.

Citing national-security concerns, it ordered the firm to prevent any non-American, including those on its own staff, from accessing its latest Fable 5 and Mythos 5 models.

In response Anthropic suspended their use altogether. The order, like the government’s clumsy attempt in late February to designate Anthropic a supply-chain risk, came out of the blue and with little explanation. It sent shockwaves through the AI world and among America’s allies.

Both sides differ about the events leading up to the intervention. Reportedly, Andy Jassy, the CEO of Amazon, an e-commerce giant that has a big stake in Anthropic, alerted the Trump administration that it had found security flaws in Fable 5, enabling it to “jailbreak”, or override, guardrails that made the model less powerful than Mythos. Posting on X, David Sacks, a White House adviser, said the government had asked Mr Amodei to “fix the jailbreak or de-deploy” Fable, adding that “Dario refused”.

A source at Anthropic said the firm was instructed to block “foreign nationals” from accessing the models because of an unspecified national-security threat. Anthropic shut them down shortly afterwards, concluding that this was the only way to comply with the directive. The company made no mention of Mr Amodei’s alleged refusal to fix the jailbreak.

It was not immediately clear on what legal grounds the government imposed the suspension. Anthropic said the models were hit by an export-control directive. Charlie Bullock, of the Institute for Law & AI, a think-tank, speculates that the government used the Commerce Department’s Export Administration Regulations, which are also used to restrict sales of AI chips to China.

But the government’s primary aim may not have been to control foreign access to frontier models. Instead, it appears to have used export controls as a convenient way to target Anthropic. If so, foreigners may be collateral damage in a power struggle between the Trump administration and a leading model-builder that President Donald Trump once described as an “out-of-control Radical Left AI company”. On June 13 Pete Hegseth, America’s Secretary of War, crowed that three months earlier he had “kicked Anthropic out” of the Pentagon.

“Every passing day proves why that was the right move.”

A lingering question is how serious the jailbreak threat really was. Anthropic said the techniques used identified only a “small number of previously known, minor vulnerabilities” that rival AI models could also uncover.

The Anthropic source added that the firm had worked with the government on pre-release testing and that officials had not previously raised any national-security concerns. Katie Moussouris, a leading authority on software vulnerabilities who reviewed the Amazon paper that sparked the furore, argues that the techniques were primarily aimed at helping cyber-defenders. Fixing the vulnerabilities, she said, would make the model less useful for cyber-defence.

The cutoff is a geopolitical watershed. It has echoes of America’s decision to restrict public-key cryptography, a technology used to secure digital communications, from the 1970s to the 1990s. Back then the government argued that cryptography was akin to a munition; one developer was investigated by the FBI for violating the International Traffic in Arms Regulations. Civil-liberties advocates eventually prevailed, securing the right to use, sell and export most encryption systems.

Encryption was a potent technology, but narrow in its application. AI is far more powerful and versatile. On June 11th Mark Warner, the vice-chair of the Senate Intelligence Committee, said that General Joshua Rudd, who leads the National Security Agency and the Pentagon’s Cyber Command, had told him that Mythos “broke into almost all of our classified systems, not in weeks, but in hours”.

Advanced AI differs from encryption in another respect, too. Whereas cryptography eventually became widely available abroad, America today enjoys a clear lead in AI. China, hobbled by American chip controls, is probably about a year behind. That advantage could become unassailable if Anthropic or other American labs crack recursive self-improvement (RSI), whereby models write better versions of themselves and thereby accelerate progress. Many insiders think that is entirely possible.

A better analogy, then, may be nuclear technology, a comparison that has inspired, fascinated and horrified AI researchers. During the second world war Britain shared its early nuclear-weapons research with America.

But in 1946, with the war over and the bomb’s awesome potential demonstrated in Japan, Congress passed the McMahon Act, severing co-operation with all foreign countries — even America’s closest allies. Full co-operation resumed much later, after Britain had already shown that it could develop its own bomb.

America’s allies are now reeling again. Many had spent months securing access to Mythos for government departments, banks and companies. Those permissions evaporated overnight.

The cut-off exempts no one, not even America’s partners in the Five Eyes intelligence alliance: Australia, Britain, Canada and New Zealand. Those countries are already grappling with the lapse of key surveillance powers that Congress recently failed to renew.

Losing access to Mythos puts them in the same boat as Russia, China or Iran. Britain’s AI Security Institute, the world’s leading body for testing and jailbreaking new models, has also been locked out.

Some foreign policymakers see the move as a wake-up call. “After a lesson this clear every nation will be asking what they need to achieve sovereignty,” argues Tom Tugendhat, a former British security minister. But Europeans and other middle powers are in a tight spot, writes Anton Leicht of the Carnegie Endowment, a think-tank.

“Do you think a Trump administration that just refused to give you access to Fable is going to let you buy enough frontier chips to train an unrestricted Fable clone yourself?”

Spy agencies are likely to regain access to Mythos, says one former British intelligence official; negotiations are already under way. Private firms may find it harder. Even so, some observers believe the American government will eventually have to relent.

“Allies can perhaps take some comfort in the fact that this is a totally untenable approach to use long term, due to the number of foreigners inside American AI companies,” says Helen Toner of Georgetown University’s Centre for Security and Emerging Technology.

Preventing foreign nationals from accessing the models is essentially equivalent to preventing any company affected from doing any further AI R&D work.

The administration’s policy may also prove ineffective. On the black market, hackers can buy American identities to gain access to AI models, as well as tools for jailbreaking them, says Cynthia Kaiser, a former official in the FBI’s cyber division. Anthropic already restricts use of its Claude model in China, yet Chinese users still access it with relative ease.

Mr Trump has pursued a bewildering approach to AI policy in recent months. He reversed most of the regulations put in place by the previous administration, which he has repeatedly mocked. He later permitted the sale of advanced AI chips to China.

In April his hands-off approach to AI safety was called into question when Anthropic produced Mythos Preview, a model the firm judged such a threat to national security that it limited its release to a small group of approved customers. On June 2 Mr Trump issued an executive order calling for a voluntary framework under which AI labs would provide the government access to their most advanced models shortly before release.

The sudden and capricious nature of the administration’s restrictions on Fable 5 and Mythos 5 is out of step with that approach—and a far cry from the consistent and transparent oversight Mr Amodei has advocated. In theory the Centre for AI Standards and Innovation, a government body that vets frontier models for dangerous capabilities, could serve as an independent arbiter in such disputes.

But in recent days the administration instructed it to stop publishing public reports. As access to advanced AI becomes a matter of national security, America’s governance of it seems increasingly opaque.