At the same G7 forum, three frontier-lab leaders reportedly called for allies to write frontier AI rules together while allied governments challenged a US block on their access to Mythos 5.

Safety assumed the participants could share the problem

Frontier AI safety was built around a common-risk premise. Labs held technical knowledge; governments held legal authority; countries might disagree over thresholds. But the model remained a shared governance problem: identify dangerous behavior, establish guardrails, compare evaluations, then coordinate a response.

That structure held while access remained secondary. A safety standard could appear neutral when participants argued over what a capable model should be permitted to do, not which participants could use it. “AI safety” named the hazard without assigning control over the machine.

But abstractions have addresses. Safety eventually arrives at a deployment decision: a model is available or blocked, a lab accepts or rejects a condition, a government permits or restricts access. Once those decisions diverge across borders, the safety framework stops operating only as a common rulebook. It becomes an allocation system.

The export decision changed the question

The shift was visible before the G7. Dario Amodei held multiple tense calls with Trump administration officials ahead of an export-controls decision. That episode placed Anthropic’s safety position inside a state decision about distribution, where arguments over guardrails could no longer be separated from arguments over who received capability.

Neither side had abandoned safety; the operating question had changed. Governments and labs were no longer asking only how to reduce the risks of frontier models. They were also asking who had the authority to decide where those models could operate—and therefore whose definition of acceptable risk would become binding.

This is the structural center of model-access geopolitics. Access is not merely another policy topic beside standards, talent, and security. It orders them. A government that can deny a model can set conditions on deployment; a lab that controls the model can negotiate those conditions; an ally without access must contest the rules from outside the capability boundary.

The G7 exposed both systems at once

At the G7, one report said Amodei, OpenAI’s Sam Altman, and Google DeepMind’s Demis Hassabis called for US-led collaboration on AI rules. A separate account of the closed-door meeting said Amodei and Hassabis sought a US-led coalition to shape rules and standards. These were reports, not announced policy; no formal coalition can be inferred from them.

What can be inferred is the proposed architecture. “US-led” assigns a center to an ostensibly collaborative system. It does not necessarily exclude allies, but it establishes an order: leadership first, coordination second. That order becomes consequential when allied governments are simultaneously challenging the US block on Mythos 5.

The allies’ objection does not disprove the coalition language. It reveals its load-bearing assumption. A standards coalition works as a partnership only while members believe they have meaningful standing over both the rules and the capability. If one government retains the decisive access gate, collaboration can remain broad in discussion while becoming narrow in operation.

This is the reversal. Safety governance began as an attempt to subject powerful models to shared constraints. It is becoming a mechanism through which control of powerful models determines who may participate in writing those constraints. The model was once the object governed by the table; access to the model now determines who gets a seat.

The labs are not simply becoming national instruments

That does not mean frontier labs have become instruments of a nationalized US agenda. OpenAI told staff that it had strongly conveyed to the US government that building AI requires global talent. The position creates a real counterforce: the state may seek tighter control over strategic capability while labs still depend on people, organizations, and markets that cross the same borders.

The collaborative language at the G7 should also be taken seriously on its own terms. A US-led coalition could be an attempt to preserve coordination rather than monopolize it. Allied resistance to the Mythos 5 restriction shows that close partners are not passive recipients of Washington’s design.

Those qualifications strengthen the structural reading. The emerging system is not a settled national bloc but a contest over which layer carries final authority: safety expertise, sovereign control, coalition legitimacy, or possession of the model. OpenAI can defend global talent and allies can resist a US block, but both return to the same bottleneck: the party controlling access can turn preferences into operating conditions; everyone else must turn theirs into an appeal.

The rulebook now follows the key

A system’s purpose is revealed by what it repeatedly does, not by the language attached to it. As frontier AI governance sorts countries and institutions by whether they may obtain the most capable models, its purpose expands beyond safety coordination to capability distribution.

That does not make the safety work insincere. Structural reversals do not require hypocrisy. They occur when a mechanism designed for one operating environment acquires a second function under new conditions, and that second function begins to dominate. Standards once organized a shared response to capability; now capability is organizing the standards.

The G7 table was set for shared blueprints. Mythos 5 revealed the keycard beneath the placemat.