The final day of this year's Group of Seven summit brought an unfamiliar sight to the negotiating room. Beside the leaders of the world's largest advanced economies sat the executives who build artificial intelligence, a marker of how quickly the technology and the people behind it have moved to the center of global politics.
Sam Altman of OpenAI and Anthropic's Dario Amodei were among more than a dozen industry figures invited to a working lunch in the French resort town of Evian on Wednesday, alongside Google DeepMind chief Demis Hassabis. The session was billed around the safe and rapid deployment of AI, with the agenda covering frontier risks, the infrastructure that powers large models, and questions of national sovereignty. France's presidential office said online child safety would also feature in the talks.
The guest list stretched well beyond the American firms. France's Arthur Mensch of Mistral, Canada's Aidan Gomez of Cohere, Italy's Uljan Sharka of Domyn and Britain's Victor Riparbelli of Synthesia were all expected, joined by Robin Rombach of Germany's Black Forest Labs. Salesforce boss Marc Benioff and Meta's Alex Wang appeared on the list too, along with the founders of India's Sarvam and Japan's Sakana.
Influence changes hands
For some analysts, the seating chart was the headline.
Jessica Brandt, who studies technology and national security at the Council on Foreign Relations, told CNBC that governments can no longer make believable promises about AI without the cooperation, and sometimes the explicit backing, of the small number of firms developing it. She described the change as a shift in who actually holds power at the top of policymaking.
Washington draws a line
That dependence was on display in a dispute that has set Anthropic against Washington. Days before the summit, Commerce Secretary Howard Lutnick wrote to Amodei to say the company's two newest models, Fable 5 and Mythos 5, would fall under export controls, barring their use by any foreign national whether inside or outside the United States. Anthropic responded by cutting off access for every customer, including its own staff abroad, and called the order a misunderstanding it hoped to resolve quickly.
The restrictions followed alarm over what the new models could do. Anthropic had launched Fable 5 only days earlier, describing it as the first of a new “Mythos-class” tier that surpassed anything the company had previously released, with a particular talent for finding software vulnerabilities. Reporting by Fortune and Axios traced the trigger to a jailbreak. A rival firm said it had bypassed the model's safeguards, and Amazon's chief executive raised the alarm, prompting the Commerce Department to act after the administration failed to persuade Anthropic to delay the release. The same letter now requires a government license before the models can move abroad or change hands inside the country, and Anthropic already runs a pre-deployment testing partnership with a standards body housed at Commerce.
Anthropic disputed the severity of the finding. The company argued the jailbreak was narrow rather than universal, unlocking a single cybersecurity capability in one specific case, and said the same trick would work on other widely available models, including OpenAI's GPT-5.5, which face no equivalent controls. Applying such a standard across the board, it warned, would freeze the release of new frontier models industry-wide.
To outside observers, the order broke assumptions that had held for years.
Emerson Brooking of the Atlantic Council said G7 governments had long planned to build their own sovereign AI capacity while still drawing on American technology, and that Washington's readiness to deny even treaty allies certain capabilities had upended that bargain. Cameron Kerry of the Brookings Institution placed the turning point earlier, calling the release of Mythos an inflection point that pushed the Trump administration toward regulating the field at all.
Competing rulebooks
The friction sits on top of a wider split in how governments treat the technology. Europe moved first with binding law, and the EU AI Act began phasing into force across 2025 and 2026, the most detailed rulebook any major economy has adopted. Washington has leaned instead on voluntary pledges and executive guidance, while Tokyo has kept its regime lighter to court developers. Evian gave these competing approaches a rare chance to meet in one room.
For the companies, the lunch offered a chance to shape policy before binding rules arrived. Brandt expects the gathering to produce a set of voluntary commitments covering youth safety and frontier risk in cyber and biology, pledges she thinks will quietly settle into the global standard. OpenAI signaled the same intent earlier this month, telling CNBC it anticipated such commitments from the talks.
The format has precedent. Under Japan's 2023 presidency, G7 leaders agreed on a set of guiding principles and a voluntary code of conduct for AI developers through what became known as the Hiroshima process, a framework Brussels welcomed as a complement to its own forthcoming law.
Anthropic, for its part, remains in talks with the administration over restoring access to its models. An official involved said the lockdown would hold until the government's national security apparatus was hardened, a step the person expected within weeks.
Comments
Join the discussion and share your perspective.