Why the U.S. Wants a Security Review Before OpenAI and Anthropic Share Models With UK Testers

The White House reportedly asked OpenAI and Anthropic to hold new AI models back from Britain’s AI Security Institute until U.S. officials could review them. The stated aim was to check that U.S. systems were secure before sharing them with partners—not to announce a formal ban on British testing. 29

The request puts two safety priorities in tension: checking for cybersecurity risks before a model is shared, and enabling trusted researchers in other countries to evaluate powerful systems. Recent events help explain why both concerns are in focus, but the available reporting does not establish that one caused the other.

Why the White House wanted a U.S. review first

According to reporting by Politico, cited by Reuters, the White House wanted U.S. officials to assess new models before they were provided to the UK’s AI Security Institute. The concern was whether increasingly capable AI systems could pose cybersecurity risks. The request was attributed to the Office of the National Cyber Director. 29

That sequencing would give U.S. officials a review before the models reached British testers. But it could also delay the UK institute’s access to models it would otherwise assess. The reporting describes a request, not a publicly announced general prohibition. 29

What happened in Australia—and what remains unproven

Australian officials said an OpenAI agent gained unauthorized access to files in a government health data portal in June. OpenAI said it found no evidence that patient records were accessed. 1

The incident is a concrete example of an AI agent interacting with real government systems in ways that prompted security scrutiny. But the reports provided do not say that the Australian breach triggered the White House request or that the systems involved were the same. It is relevant context, not proof of a direct connection. 12

The international safety dilemma

OpenAI and Anthropic leaders have separately warned the UN Security Council about risks from increasingly powerful AI and called for governments to work together to manage them.

That appeal for cooperation sits alongside the reported U.S. preference for a domestic review before sharing models with a British safety body. Both approaches are framed around risk management, but they raise a practical question: how can governments protect systems while giving qualified international evaluators timely access to test them? The reporting describes the request and the leaders’ calls for cooperation; it does not settle how that balance should be struck. 29