G7 discussions over access to cutting-edge U.S. artificial intelligence models have placed Washington’s emerging model-control regime at the center of technology policy, cybersecurity strategy and transatlantic industrial competition.

At the summit in Evian-les-Bains, France, leaders discussed a plan that could grant select “trusted partners” access to advanced AI models developed by U.S. companies such as Anthropic, Reuters reported, citing three diplomatic sources. The proposal would create a potential path around recent U.S. restrictions on non-American use of certain frontier systems and could cover either governments or private companies, depending on how the policy is ultimately designed.

The talks follow a sharp escalation in U.S. scrutiny of frontier AI deployment. Anthropic disabled access to Fable 5 and Mythos 5, described as its most advanced models, after President Donald Trump ordered the company to block foreign nationals from using them on national security grounds, according to Reuters. The move turned what had been a largely technical question of model safety into a diplomatic and commercial issue for U.S. allies that rely on American AI systems, cloud infrastructure and cybersecurity vendors.

The G7 discussions were held mainly with U.S. Commerce Secretary Howard Lutnick on the sidelines of the summit’s opening dinner, Reuters reported. The proposed trusted-partner framework would be intended to give allied countries access to powerful models for defensive purposes, including cybersecurity work against rivals such as China. The White House said the administration had an open line of communication with allies and remained committed to addressing national security concerns tied to Anthropic’s model.

The issue is especially sensitive because the technology at the center of the debate is dual-use. Anthropic’s Mythos model was designed to identify flaws in computer code, a function that can help banks, hospitals, utilities and communications networks detect vulnerabilities before hostile actors exploit them. But the same capabilities could also accelerate offensive cyber operations if used by criminals or state-linked hackers. Reuters reported that cybersecurity experts believe Mythos could turbocharge attacks on bank technology systems, while the European Union has sought access to study the model’s implications.

Before the U.S. order, Anthropic had given select organizations in more than 15 countries access to Mythos so they could scan systems for vulnerabilities, according to the company statement cited by Reuters. The organizations included entities in healthcare, communications, power and water. That prior deployment illustrates why allied governments are pressing for a more predictable access channel: the model may be a national security risk, but it may also be a tool for protecting national infrastructure.

The commercial implications extend beyond Anthropic. If the United States formalizes a trusted-partner regime for frontier AI, model access could become a controlled strategic resource, similar in policy logic to advanced semiconductors, cloud computing capacity and cybersecurity tools. That would affect not only model developers but also cloud providers, enterprise software companies, financial institutions, critical-infrastructure operators and government contractors that depend on access to the strongest AI systems for code review, threat detection, fraud monitoring and automated operations.

The White House had already moved in that direction before the G7 summit. A June 2 executive order on advanced AI innovation and security directed senior U.S. officials to develop a classified benchmarking process for assessing advanced cyber capabilities and determining when a system should be treated as a “covered frontier model.” The order also called for a voluntary framework under which AI developers could give the federal government access to such models before wider release, subject to confidentiality, cybersecurity, insider-risk, intellectual-property and nondisclosure protections.

Crucially, the executive order also contemplated collaboration between the federal government and AI developers to select trusted partners that would receive early access to covered frontier models in order to promote secure innovation and strengthen critical-infrastructure cybersecurity. That language now appears central to the G7 debate. Allied governments are not simply asking for looser controls; they are seeking a mechanism that would allow access under defined security conditions while preserving Washington’s ability to manage proliferation risk.

For Europe, the question lands amid a broader push for digital sovereignty. European Commission President Ursula von der Leyen said on Wednesday that it was in the mutual interest of the United States and the European Union for European companies and citizens to be able to safely use the best AI models. Speaking at a G7 lunch with leaders and technology executives, she framed the issue around trust, safety and interdependence, noting that the U.S. and EU already use one another’s trusted technologies and have interconnected financial systems.

Her comments reflected the central European dilemma. The EU wants access to top U.S. AI systems because they may be necessary for competitiveness and cyber defense. At the same time, sudden U.S. restrictions on Anthropic’s models reinforce European concerns that dependence on American technology can become a strategic vulnerability. The more Washington treats frontier models as controlled assets, the stronger the incentive becomes for Europe to support domestic AI labs, sovereign cloud arrangements and local computing infrastructure.

The G7 format gives the dispute a broader geopolitical frame. The group includes the United States, France, Germany, Italy, Japan, Canada and the United Kingdom, with the European Union also participating. These economies are closely aligned on many security questions, yet they have different industrial positions in AI. The United States hosts the most influential frontier model companies and hyperscale cloud platforms. Europe has important AI startups and regulatory influence but remains more dependent on imported compute capacity and U.S.-controlled platforms. Japan, Canada and the United Kingdom are also trying to balance access, domestic capability and security oversight.

The involvement of AI executives further underscores that any access framework would need industry implementation. Reuters reported that executives from Anthropic, OpenAI and Google were expected to attend a working lunch on technology issues including regulation, AI infrastructure and networks. The Associated Press reported that OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis and Anthropic CEO Dario Amodei were among the AI leaders due to join G7 discussions focused on safe, rapid and effective deployment of the technology.

For model companies, the policy shift presents both risk and leverage. On one hand, sudden restrictions can disrupt product launches, enterprise contracts and international revenue plans. On the other, being designated as a developer of strategically important models may deepen ties with governments and critical-infrastructure customers. Companies able to satisfy security, audit, access-control and monitoring requirements may gain an advantage in serving regulated sectors. Smaller AI providers, especially those outside the United States, may use the episode to pitch sovereign or open-weight alternatives to customers worried about U.S. policy intervention.

Cloud providers could also be affected. Advanced AI access is increasingly tied to where models are hosted, how inference is logged, what user categories are permitted, and whether sensitive workloads can be isolated by geography or legal jurisdiction. A trusted-partner regime may require cloud and AI vendors to build more granular controls around nationality, residency, customer type, use case and data handling. That would raise compliance costs but could also create demand for premium secure AI environments designed for governments, banks and critical-infrastructure operators.

The cybersecurity industry may see the most immediate operational impact. Defensive AI tools that can scan large codebases, identify vulnerabilities, prioritize patching and test system resilience are becoming more valuable as organizations face increasingly automated threats. If the most capable models are restricted, security teams outside the United States could face a capability gap unless trusted access is approved. Conversely, if access is widened too quickly, governments risk putting powerful vulnerability-discovery systems into environments where misuse, leakage or compromised credentials could create new attack surfaces.

The policy challenge is therefore not simply whether allies should receive access, but how such access should be verified and monitored. A workable framework would likely need clear eligibility rules, secure deployment environments, logging and auditing requirements, incident reporting, user vetting, export-control compliance and limits on model fine-tuning or redistribution. It would also need to define whether trusted partners are sovereign governments, licensed companies, critical-infrastructure operators, research institutions or some combination of those groups.

Washington’s balancing act is complicated by the competitive pace of AI development. Overly restrictive controls could push customers toward rival systems, including European providers, open models or Chinese alternatives. A looser regime could undercut the rationale for restricting foreign access in the first place. The United States also has to manage the expectations of allies that contribute to its broader security network and may argue that defensive access to AI models should be part of intelligence, cybersecurity and industrial cooperation.

The G7 discussions did not produce a final public agreement on a trusted-partner model-access framework. But the fact that the issue reached leader-level talks so soon after the Anthropic restrictions shows how rapidly frontier AI has moved into the same policy category as chips, critical minerals, telecom infrastructure and cyber weapons. Access to advanced models is becoming a matter of alliance management, not only product distribution.

For investors and enterprise technology buyers, the near-term signal is that frontier AI governance is becoming more specific and more operational. Future model launches may be influenced by classified benchmarks, government review, critical-infrastructure priorities and trusted-user designations. The winners may be companies that can combine frontier performance with demonstrable security controls and government-grade compliance. The losers may be firms and customers that assumed advanced AI access would remain globally uniform and commercially determined.

The G7 debate also suggests that the next phase of AI regulation may be less about broad declarations of safety and more about access architecture: who gets the model, where it runs, what it can do, and how governments verify that use remains defensive. As the United States tries to preserve AI leadership while keeping allies inside its technology perimeter, the trusted-partner question is likely to become a recurring test of whether frontier AI can be governed without fragmenting the market it helped create.