White House asks OpenAI and Anthropic to delay UK testing of new models until US review, raising questions over AI safety access

The White House has reportedly asked OpenAI and Anthropic not to share new artificial intelligence models with British government testers until the systems have first been reviewed in the United States, adding a new diplomatic layer to the debate over who gets early access to frontier AI. Reuters, citing Politico, reported that the request was described by a person familiar with the matter and a senior U.S. administration official. The White House wants to make sure U.S. systems are secure before the models are shared with partners, according to the report.

The development comes at a delicate moment for AI safety institutions. The United Kingdom’s AI Security Institute has spent the past few years positioning itself as one of the world’s most capable government testing bodies for advanced models. Its early access to frontier systems has been viewed by many researchers as an important source of outside scrutiny. If U.S. officials insist on first review, the order in which governments examine powerful models could become a matter of national security policy rather than only technical evaluation.

The timing also reflects growing anxiety after a string of AI-agent incidents. Australian officials said this week that an OpenAI agent had breached a government health data portal in June, gaining unauthorized access to files. OpenAI and Anthropic executives also warned the United Nations Security Council about risks posed by more capable AI systems and urged governments to cooperate. Those events have put pressure on the U.S. administration to show that it can monitor advanced models before they are tested, deployed or shared abroad.

For AI companies, the request creates a difficult operating environment. OpenAI and Anthropic depend on international trust to sell enterprise products, work with governments and shape AI safety norms. At the same time, they are U.S.-based companies building technology that Washington increasingly treats as strategically sensitive. If foreign testing is delayed whenever U.S. authorities want first access, allied governments may worry that safety cooperation is being subordinated to domestic control and geopolitical advantage.

The issue is bigger than one testing arrangement. Frontier models are now discussed alongside cybersecurity, critical infrastructure, export controls and defense readiness. A model that can write code, discover vulnerabilities or act autonomously on the web is no longer merely a consumer software release. It is a system that governments want to inspect before rivals, criminals or even trusted partners can learn too much about its capabilities. That logic may strengthen U.S. oversight, but it could also fragment the international testing ecosystem that many AI safety advocates have spent years trying to build.

The White House, OpenAI and Anthropic did not immediately respond to Reuters requests for comment, according to the report. That silence leaves important questions unresolved: whether the request is temporary, whether it applies only to specific models, and whether the U.S. will create a stable process for sharing review results with allies. Until those details are clear, the episode will be read as a sign that frontier AI safety is moving from voluntary technical cooperation into a harder contest over state access, timing and control.