Washington Reverses Export Curbs on Anthropic's Claude After Security Review

A Swift About-Face

The Commerce Department has cleared Anthropic's two newest large language models for export, ending a brief but consequential restriction that had blocked international distribution and raised questions about how Washington evaluates frontier AI capabilities. Fable 5 is now available worldwide, while Mythos 5 has been accessible to US organizations since late June. The reversal comes after Anthropic worked directly with federal agencies to demonstrate risk mitigation measures around the models, which the administration had initially deemed sensitive enough to warrant export controls.

Commerce Secretary Howard Lutnick informed Anthropic in writing that the company no longer requires licenses for cross-border transfers of either model. His letter acknowledged that Anthropic had coordinated closely with government officials to address the concerns that triggered the original hold. The episode marks one of the first times a sitting administration has invoked national security authority to restrict an AI model's distribution, then lifted those curbs after a technical review process.

The Glasswing Pathway

Mythos 5 remains under a tiered access framework. While US-based entities regained full use last week, Anthropic is now negotiating with regulators to extend availability through the Glasswing program, a government-endorsed initiative that grants vetted cybersecurity researchers at approved organizations access to high-capability models for defensive research. The program was designed to balance the need for cutting-edge tools in threat intelligence with the risk that the same tools could be weaponized if widely released.

Glasswing participation requires organizations to pass a screening process and commit to use restrictions. Anthropic has indicated it expects to expand the roster of domestic and international partners eligible for Mythos access over the coming months, though the company has not disclosed timelines or named prospective participants. The framework reflects a broader policy trend: as models grow more capable, governments are experimenting with conditional release mechanisms rather than binary approve-or-block decisions.

What Triggered the Hold

The initial export restriction arrived without public warning in early June. Administration officials cited unspecified national security concerns, a designation that can encompass everything from dual-use capabilities in cyber operations to risks of foreign adversaries fine-tuning models for disinformation or surveillance. Anthropic did not publicly contest the hold, instead entering a series of technical briefings with Commerce, the National Security Council, and the AI Safety Institute.

Industry observers noted that the move came at a moment of heightened scrutiny around model capabilities. Both Fable 5 and Mythos 5 represent generational leaps in reasoning and code generation, domains where the line between commercial utility and security risk is especially thin. The fact that the curbs were lifted relatively quickly suggests the government found Anthropic's safeguards credible, but the episode has left developers across the sector recalibrating their assumptions about regulatory timelines.

Implications for the Sector

The three-week saga offers a preview of the regulatory environment that frontier labs will navigate as model capabilities continue to scale. Export controls, traditionally reserved for semiconductors and encryption, are now being adapted for software artifacts that can be copied and deployed anywhere. The speed of this cycle - from restriction to clearance - suggests that agencies are still refining their evaluation criteria and that close collaboration with developers can accelerate reviews.

For Anthropic, the outcome is a validation of its voluntary safety testing regime, which includes pre-deployment evaluations for biological, chemical, and cyber risks. The company has positioned itself as a partner to regulators, a strategy that appears to have paid dividends in this instance. Other labs, particularly those with less established government relationships, may face longer holds or more prescriptive conditions.

The Glasswing model, meanwhile, could become a template for controlled releases. By creating a middle tier between full public availability and outright prohibition, policymakers gain a mechanism to monitor real-world use while limiting exposure. Whether that approach scales as the number of frontier models grows - and as international coordination lags - remains an open question.

What Comes Next

Anthropic has confirmed that both models are now integrated into its commercial API and enterprise offerings. Fable 5, the more widely accessible of the two, is already seeing uptake in customer service automation, legal research, and software development workflows. Mythos 5, with its more specialized capabilities, will likely remain in narrower circulation even as Glasswing expands.

The episode has also prompted calls for clearer rulemaking. Industry groups have urged the Commerce Department to publish criteria for when export controls apply to AI models, arguing that ad hoc decisions create uncertainty for investors and deployment timelines. So far, the administration has signaled that it prefers case-by-case reviews, a stance that gives regulators flexibility but leaves developers guessing.

At DailyTechWire, we've tracked the evolution of AI export policy across Seoul, Singapore, and Brussels, and the Anthropic case underscores a pattern: governments are willing to experiment with novel regulatory instruments, but they are doing so faster than they are building the institutional capacity to administer them consistently. The result is a patchwork system in which a company's ability to navigate bureaucracy can matter as much as the technical merits of its safety work.

For now, Anthropic has cleared the hurdle. The broader question - how many more hurdles lie ahead, and for whom - is only beginning to take shape.