The Trump administration has directed both OpenAI and Anthropic to restrict access to their newest AI models to a curated set of government-vetted customers as part of an ongoing national security cybersecurity review — marking the first time OpenAI models have been placed under government-mandated access controls and extending restrictions on Anthropic beyond its previously announced export-control actions.
OpenAI’s GPT-5.6 Becomes First OpenAI Model Subject to Government Access Controls
OpenAI has restricted GPT-5.6 — released in Sol, Terra, and Luna variants — to a small number of companies approved through a government engagement process. The restriction represents a new category of US government intervention in commercial AI availability: previous AI-related government actions had focused on export controls governing who could receive US AI models internationally. The current cybersecurity review extends that concept to domestic commercial availability, limiting which US customers can access a new model regardless of export jurisdiction.
The Trump administration’s direction to OpenAI comes after the government had separately moved against Anthropic in prior weeks. Anthropic had previously suspended global access to its Fable 5 model under export control provisions. The current review extends to Anthropic’s full model access program, not just Fable 5, and is framed as a security assessment distinct from the export control mechanism.
The Trigger: Classified System Vulnerabilities and a New Cybersecurity Review Frame
The administration has framed the current restrictions as a security assessment of AI model capabilities in the context of national security, following the discovery by Anthropic’s Mythos model of vulnerabilities in classified US government systems — a disclosure that occurred in late June. That event appears to have accelerated the government’s posture toward treating frontier AI model capabilities as a security consideration requiring review before general commercial release.
The distinction between the current review and prior export controls is deliberate. Export controls govern which countries and entities can receive US-origin AI models across national boundaries. The cybersecurity review, as framed, is an assessment of what capabilities should be commercially available to any customer while a security analysis of those capabilities is ongoing.
Industry Pushback: More Than 100 Cybersecurity Executives Oppose the Restrictions
The access restrictions have drawn opposition from US technology industry leaders. More than 100 cybersecurity executives have signed a letter urging the administration to ease AI model restrictions, arguing that the controls harm US competitiveness relative to non-US AI providers operating without equivalent government-mandated access limits.
The competitive dimension of the concern is direct: if US AI companies are required to restrict their newest models to vetted customers while foreign AI providers face no such constraint, the effect is to reduce the availability of frontier US AI tools to developers, businesses, and security practitioners who might otherwise access them. The signatories argue that this dynamic advantages competitors outside US jurisdiction.
What the Review Means for OpenAI and Anthropic’s Commercial Model Programs
For OpenAI, the GPT-5.6 restriction is a departure from its standard commercial release model, under which new API models become available to paying customers through the standard OpenAI platform upon release. The requirement to route access through a government-vetted approval list introduces a layer of administrative gatekeeping that OpenAI has not operated under for previous model releases.
For Anthropic, the extension of the review to its full model access program broadens the government’s involvement beyond the Fable 5 export control situation. The full program scope means the review is not limited to a specific model but applies to Anthropic’s pipeline of available models during the assessment period.
The precise timeline for the cybersecurity review has not been publicly disclosed. Both companies are operating under the access constraints in parallel, which creates a situation in which the US government is simultaneously restricting two of the country’s leading AI providers’ newest models from general commercial availability.
The restrictions represent a significant moment in the relationship between US AI companies and the federal government: a shift from the government as a customer and regulator of AI applications to the government as a gatekeeper for which customers can access the most capable AI models at the time of their release. How the review resolves — and whether it establishes a precedent for future model release cycles — will shape the commercial operating environment for US AI developers going forward.
