U.S. Export Controls Altered AI Model Access
New government directives mandated restricted deployment of advanced AI models by Anthropic and OpenAI starting in June 2026.
Updated on Sept. 24, 2026 in Artificial Intelligence
Live Poll
Should the U.S. prioritize strict national AI controls over international safety testing collaboration?
On June 12, 2026, Anthropic suspended access to its Fable 5 and Mythos 5 models for all foreign nationals following a U.S. government export control directive. Simultaneously, OpenAI began a restricted preview of GPT-5.6 Sol on June 26, 2026, with participants disclosed to U.S. authorities.
Why it matters
These restrictions signal an intensifying government focus on the national security risks posed by advanced language models, particularly regarding potential jailbreak vulnerabilities. The move effectively bifurcates global research access, forcing AI developers to tailor model availability based on geopolitical and regulatory boundaries.
Anthropic employs separate safeguard configurations labeled Fable and Mythos for the same base model architecture. While Fable and Mythos underwent thousands of hours of safety testing, Mythos 5.1 is now limited to selected U.S. organizations to comply with the directive.
The players
Anthropic
An AI safety and research company focused on building steerable, large-scale language models with a heavy emphasis on constitutional safeguards.
OpenAI
A prominent developer of foundation models, currently focused on scaling GPT-series architecture for both commercial and research applications.
UK AI Security Institute
A national body tasked with evaluating the safety and potential risks of emerging artificial intelligence models.
The details
The restriction involves specific safeguard mechanisms designed to prevent jailbreak attacks—attempts to bypass AI safety filters to generate prohibited content. OpenAI manages its deployment by sharing the identity of preview participants with the U.S. government, ensuring oversight of who accesses the GPT-5.6 Sol architecture. Anthropic maintains its models as a singular underlying system, but applies distinct filter layers to differentiate between its Fable and Mythos versions in response to export control requirements.
Timeline
June 12, 2026: Anthropic suspended access for foreign nationals after receiving a government directive.
June 26, 2026: OpenAI initiated a limited preview for GPT-5.6 Sol with trusted partners.
September 16, 2026: A UK parliamentary question was submitted regarding Anthropic's cooperation with safety testing.
September 30, 2026: The UK government is expected to respond to the parliamentary inquiry.
The Tech Race
These directives mark a definitive pivot toward treating high-performance AI weights as critical dual-use infrastructure rather than open research assets. The U.S. regulatory approach now sets a standard that competing global jurisdictions, such as the UK, are being forced to navigate through formal legislative inquiry.
Developers and researchers outside of the United States face significant hurdles in accessing the latest model versions for safety evaluation or integration. Users are now subject to regional model availability, where access is increasingly gated by institutional affiliation and compliance with U.S. export guidelines.
The takeaway
The era of uniform global access to the most capable foundation models has ended as governments treat these systems as strategic national assets. Watch for the September 30 response from the UK parliament to gauge how international safety testing protocols will adapt to these restrictive U.S. mandates.
What happens next
The UK government is scheduled to provide a formal response to the parliamentary inquiry regarding Anthropic's safety testing access by September 30, 2026.
Further reading
For more on the regulation of foundation models, visit Artificial Intelligence.
Live Poll
Should the U.S. prioritize strict national AI controls over international safety testing collaboration?
