G7 leaders have begun negotiating a “trusted partners” framework that would restore select allied nations and companies’ access to the most advanced U.S. AI models — a direct diplomatic response to President Donald Trump’s order that shut non-Americans out of Anthropic’s frontier systems.
According to three diplomatic sources cited by Reuters, the discussions took place primarily on the sidelines of the G7 summit’s opening dinner on Monday in Evian-les-Bains, France, involving country representatives and U.S. Commerce Secretary Howard Lutnick. The talks were described as ongoing, with a second source saying “trusted partners” could encompass both countries and companies. A Trump White House official confirmed the administration has “an open line of communication with our allies” on national security concerns related to Anthropic’s models.
What’s New
The G7 access discussions were triggered directly by Anthropic’s shutdown of its two most advanced AI systems. On Friday, June 13, Anthropic disabled global access to Fable 5 and Mythos 5 under a U.S. government order, cutting off all foreign nationals from the models following President Trump’s directive citing national security. The move was abrupt: prior to the order, Anthropic had provided Mythos access to organizations in more than 15 countries across healthcare, communications, power, and water sectors, according to a company statement.
Mythos 5 is particularly sensitive. Cybersecurity experts believe the model — designed to detect flaws in computer code — could dramatically accelerate attacks on critical financial infrastructure if it fell into adversarial hands. That dual-use risk is central to the White House’s reasoning for restricting access. Yet the same capability is precisely what allied governments want: the European Union has specifically sought access to Mythos to study its security implications, according to Reuters.
There is a notable strategic tension embedded in Washington’s position that the G7 talks have begun to surface. By pulling Mythos access from allied cybersecurity organizations, the U.S. simultaneously removed a defensive tool from partners it relies on for collective security — while doing nothing to prevent adversarial states from independently developing equivalent capabilities. The “trusted partners” framework, if it materializes, would represent an acknowledgment that blanket export controls on AI can undercut alliance cohesion as much as they protect national security. That is a concession with significant long-term policy implications, as cybersecurity leaders have already argued publicly in demands to lift the Anthropic restrictions.
Anthropic, OpenAI, and Google executives are expected to attend a working lunch at the summit on Wednesday, Reuters previously reported, where technology regulation, AI infrastructure, and networks are on the agenda. Anthropic’s spokesperson did not immediately respond to requests for comment at the time of publication.
How the “Trusted Partners” Scheme Compares to Existing AI Access Frameworks
The proposed G7 framework is not without precedent, but it would represent a meaningful departure from how AI access has been governed to date. The table below maps it against the two most relevant existing approaches.
| Framework | Who Decides Access | Scope | Legal Basis | Key Limitation |
|---|---|---|---|---|
| Trump Executive Order (current) | U.S. President / Commerce Dept. | Blanket ban on foreign nationals for named models | National security executive authority | No ally exemption; cuts off friendly governments equally |
| Proposed G7 “Trusted Partners” Scheme | U.S. Commerce Secretary + G7 counterparts | Selective access for vetted countries or companies | To be negotiated; no formal agreement yet | Criteria for “trusted” status undefined; implementation timeline unclear |
| Anthropic’s Pre-Ban Model | Anthropic (company-level vetting) | 15+ countries; critical infrastructure sectors only | Commercial terms + voluntary security review | No government oversight; revocable unilaterally by U.S. order |
The comparison underscores that any durable “trusted partners” arrangement will require a formal intergovernmental legal basis — something that does not yet exist. Without it, access granted today could be revoked by the next executive order just as rapidly as it was shut down on Friday.
The stakes extend well beyond Anthropic. AI is already reshaping the vulnerability exploitation race, compressing the window between a flaw’s discovery and its weaponization. Allies that lose access to frontier defensive AI tools — even temporarily — face a measurable degradation in their ability to keep pace. That urgency is what pushed G7 representatives to raise the issue within 72 hours of the Anthropic shutdown, rather than waiting for formal diplomatic channels.
Google and OpenAI, whose executives are also attending Wednesday’s working lunch, develop models at comparable or greater capability levels. Neither company’s models have been subject to the same access restrictions as Anthropic’s — a disparity that, according to the source article, has not been publicly explained by the administration. Whether the “trusted partners” framework, if formalized, would extend to models from those companies is also unresolved.
How Serious Players Should Respond
For allied governments, the immediate priority is to formalize what is currently an informal dinner conversation. A voluntary framework negotiated on summit sidelines offers no legal durability; the EU, UK, Japan, and Canada should press Washington for a binding bilateral or multilateral instrument that codifies access rights, security review standards, and revocation procedures before the next executive order renders the point moot. The EU’s specific interest in studying Mythos 5 for its security implications is a natural anchor for that negotiation.
For technology executives — particularly at Anthropic, OpenAI, and Google — the G7 talks signal that government access policy is now a first-order strategic variable, not a compliance footnote. Companies that invest in transparent, auditable access-control architecture are better positioned to satisfy both Washington’s national security requirements and allies’ demands for access. The abruptness of Friday’s shutdown illustrates what happens when that architecture does not exist in advance. Satya Nadella’s broader warning about a handful of AI winners hollowing out entire industries takes on a geopolitical dimension here: concentrated AI capability, subject to unilateral government control, is a systemic fragility for the entire allied technology ecosystem.
For regulators and policymakers watching from outside the G7, the Evian talks establish a template — and a precedent. If the “trusted partners” scheme moves forward, it will effectively create a tiered global AI access regime: U.S. insiders, vetted G7 allies, and everyone else. Understanding which tier your jurisdiction falls into — and how to move up — is now a live policy question with direct implications for national AI strategy, cybersecurity posture, and economic competitiveness.
