When the US government blocked Anthropic from exporting its newest Fable 5 and Mythos 5 models last week, the immediate story was a security dispute between a company and the state. A week on, that dispute has grown a second life as a diplomatic one, and the questions it is raising are larger than any single jailbreak. The trigger was almost mundane: Amazon flagged to the White House that safety guardrails in the models could be bypassed, and within days the Commerce Department cut off foreign access on national security grounds. The models remain frozen. What has not stayed still is everyone watching.
At the G7, according to reporting collected by Silicon Canals, French President Emmanuel Macron and Indian Prime Minister Narendra Modi pressed Donald Trump on the obvious implication. If a country builds its workflows, its services and parts of its critical infrastructure on American models, what happens when Washington can switch those models off overnight, for reasons it may never disclose? Macron's argument was commercial rather than ideological: a kill switch known to exist devalues the product attached to it, and nobody will standardize on American AI if it can vanish without warning. Modi's was about sovereignty. Both were describing US AI the way energy ministers describe a pipeline that runs through someone else's territory.
South Korea offers the case study in miniature. On 18 June its science ministry signed a memorandum with Anthropic to deepen cooperation on AI safety and cybersecurity, a follow-up to talks between Deputy Prime Minister Bae Kyung-hoon and Dario Amodei earlier in the year. Yet the ink arrives shadowed. Only this month Anthropic had extended Mythos access to Korean bodies including the country's internet security agency, Samsung and SK Telecom, and the US export block has now thrown that access, and the broader partnership, into doubt. A signing ceremony and a freeze, in the space of weeks, tell the story better than any white paper.
The talks that matter most are quieter. After negotiations nearly broke down when Anthropic refused to pull Fable from the market, the two sides have, per Politico's account relayed in Indian coverage, gone back to the table to draft a shared framework for judging how serious a future AI security flaw really is and what government action it warrants. The White House has made its price clear. As Wired reported, officials told Anthropic to stop arguing that the risks are overblown and start patching the jailbreaks, pointing out that neither the NSA nor Commerce's new standards center has the staff to police every loophole in every model that ships.
There is an awkward technical footnote under all of this. Independent cybersecurity researchers have noted that the capabilities the government cites are also present in freely available models, including from OpenAI, and a growing number of computer scientists argue that durable guardrails against a determined attacker may be mathematically out of reach. If that is right, the framework being negotiated is less a fix than a treaty about how to live with an unfixable problem.
The deeper effect is already visible. Cohere's Aidan Gomez says the episode confirms what his company has long argued, that leaning on a handful of US labs is a strategic risk, and the scramble for sovereign AI capacity is accelerating from Paris to Bangalore. G7 leaders floated a trusted partners scheme to guarantee access for allies. Whether Washington signs on will decide something simple but large: whether American AI stays the global default, or becomes another system other powers feel compelled to route around.