How the EU Finally Got Access to Anthropic's Most Powerful AI — and What the G7 Trusted Partners Framework Means for Global Cyber Defense

Here is what happened and what it means.

The Model That Triggered a Global Emergency

Public reporting in April 2026 described Claude Mythos Preview as a frontier AI cybersecurity model that Anthropic considered too powerful for general public release because of its ability to find and exploit software vulnerabilities . The UK's AI Safety Institute later reported that, in controlled evaluations, Mythos Preview could execute multi-stage attacks on vulnerable networks and autonomously discover and exploit vulnerabilities — tasks that would normally take human professionals days of work .

This triggered significant international concern around AI-enabled cyber risk. Rather than releasing the model broadly, Anthropic limited access through Project Glasswing, a controlled-access initiative intended to help selected organizations improve resilience against the kinds of vulnerabilities Mythos can identify .

Project Glasswing: From 50 to 200 Partners

Anthropic expanded Project Glasswing by granting roughly 150 new organizations across more than 15 countries access to Claude Mythos Preview, bringing total partners to around 200 . The new cohort intentionally filled sectors underrepresented in the original pilot, including power, water, healthcare, communications, and other critical infrastructure areas where major vulnerabilities could have broad public impact .

The expansion came after an initial cohort of about 50 partners — including Amazon, Apple, and Microsoft — had already used Mythos to surface more than 10,000 high- or critical-severity software flaws since the program launched in early April .

The San Francisco Meeting: How ENISA Got In

After weeks of contentious negotiations, Anthropic agreed to give ENISA access to Mythos through Project Glasswing, making it the first EU institution reported to receive access . Reports state that Anthropic invited the European Commission to facilitate ENISA's access to the cybersecurity-focused Mythos model .

The decision ended a period of negotiation over whether EU authorities would be able to evaluate one of the most advanced U.S.-developed AI security tools . European Commission spokesperson Thomas Regnier confirmed the bloc had "several productive meetings" with Anthropic and welcomed the developments .

However, the EU still needs to establish a mechanism with appropriate security safeguards before access can be implemented; an ENISA official confirmed the agency does not currently have direct access .

G7 'Trusted Partners': A Framework for AI Access

At the June 15–17 G7 summit in Évian-les-Bains, leaders discussed a plan to grant select "trusted partners" access to advanced AI models from U.S. firms such as Anthropic, according to three diplomatic sources cited by Reuters . The talks, held on the sidelines of the opening dinner, involved U.S. Commerce Secretary Howard Lutnick and country representatives .

The discussions followed the Trump administration's export control directive, which prompted Anthropic on June 13 to abruptly disable access to its most capable models — Fable 5 and Mythos 5 — for all customers globally .

French President Emmanuel Macron expressed optimism that progress would be made in coming weeks on broadening access to leading U.S. AI models . The framework being discussed could potentially allow either countries or companies to receive privileged access based on shared security standards and trust criteria .

Europe's AI Sovereignty Challenge

Despite securing ENISA access to Mythos through Project Glasswing, Europe's position remains dependent on access to a U.S.-controlled model rather than ownership of the underlying technology . Anthropic has described Mythos as too powerful for general public release and has limited access to carefully selected parties, which means European access appears to be governed rather than sovereign .

The provided sources do not identify a European equivalent to Mythos-level autonomous cyber capabilities. They also do not specify whether ENISA's access covers full model capabilities or a limited Project Glasswing access tier, so that point remains open unless confirmed by additional reporting .

The broader picture is therefore best described as managed access, not full AI sovereignty: European institutions may be gaining evaluation or defensive use of Mythos, but the model itself remains controlled by Anthropic .

What Comes Next

Two parallel tracks are now in motion. In the near term, the EU must build the security infrastructure to actually use Mythos — a process that could take weeks or months . In parallel, G7 diplomats are working to formalize the trusted partners framework, with Macron predicting progress "in coming weeks" .

The outcome may determine not just who can use the world's most advanced AI models, but whether the current U.S.-centric model of AI governance gives way to a more multilateral system — or hardens into an exclusive club with America at the center.