An Amazon research team reported a method that appeared capable of bypassing safeguards in Anthropic’s Fable 5 model. A US export directive followed, temporarily restricting access while Anthropic investigated and strengthened the safeguards. Anthropic restored access on 1 July 2026. The episode also sharpened a question raised at the G7: what happens when an economy depends on AI infrastructure that another government can restrict?
At the G7 Summit, French President Emmanuel Macron and Indian Prime Minister Narendra Modi pressed it directly: what happens to a country’s economy when the AI models it depends on can be switched off by a foreign government?
The trigger: Anthropic’s Mythos 5 and Fable 5 export block
Days before the summit, the Trump administration temporarily restricted exports of Anthropic’s newest models on national-security grounds. Anthropic’s later account said the directive followed a report in which Amazon researchers described a method for bypassing some Fable 5 safeguards.
Anthropic said it investigated the report, introduced additional safeguards and restored access to Fable 5 and Mythos 5 on 1 July. The temporary restriction nevertheless demonstrated that access to frontier models can be affected by government action with little notice.
For governments and enterprises that had built workflows on top of those models, the lesson was immediate: access can be revoked overnight, for reasons that may never be disclosed.
Macron and Modi articulate the structural risk
Macron warned that the U.S. could revoke access to AI infrastructure overnight, harming European customers’ economies and damaging American AI firms themselves. Modi, sitting across the table, said democratic nations must have unfettered access to top AI models to protect critical infrastructure.
The framing matters. Two heads of state from the world’s largest democracies are describing American AI infrastructure the way energy ministers describe a pipeline running through hostile territory: useful, indispensable, and a strategic liability.
The ‘trusted partners’ workaround
G7 leaders discussed creating a trusted partners scheme that would grant non-U.S. nations and companies access to advanced models from firms like Anthropic and OpenAI, conditional on using them to harden defences against rivals such as China. Macron argued Washington should back the framework and broaden Mythos access, on the simple commercial logic that nobody will buy U.S. AI if it can disappear overnight.
The unresolved question is scope. It is unclear whether a startup in Paris or Bangalore whose product breaks without warning would qualify for protection, or whether the scheme covers only sovereign customers and large enterprises.
The sovereign AI scramble accelerates
Aidan Gomez, co-founder and CEO of Canadian enterprise AI firm Cohere, said the episode confirmed what his company had long argued: that dependence on a small handful of big tech companies is dangerous to the resilience of companies and democratic nations alike. He described digital sovereignty as a question of control over the foundational technology that will shape economic security and national sovereignty for decades to come. That message is landing. The Anthropic episode strengthened existing arguments for sovereign AI capacity, and Silicon Canals has previously examined how Africa’s $60B AI sovereignty plan still runs through 12,000 Nvidia GPUs and Big Tech infrastructure, a contradiction now visible at the G7 level.
What the episode reveals about the AI export stack
The structural picture is straightforward. American frontier labs are subject to U.S. export controls. In this case, the controls followed a report from Amazon researchers concerning a possible safeguard bypass, while competing models remained available. International customers absorb the political risk. American firms absorb the lost revenue. The U.S. government retains the kill switch.
Macron’s point to Trump was commercial rather than ideological: a kill switch known to exist devalues the product attached to it. The G7 discussions are an attempt to negotiate a contractual ceiling on that switch. Whether Washington agrees will determine whether American AI remains a global default or becomes, like American payment rails before it, a system other powers feel compelled to route around.
