The U.S. Office of the National Cyber Director has requested that OpenAI and Anthropic submit new AI models to U.S. agencies for evaluation before providing them to the U.K.’s AI Safety Institute (AISI), according to a report from Politico. The directive places American frontier AI developers under pressure to prioritize domestic oversight over international testing partnerships. In response, Anthropic has restricted access to its Claude Mythos 5.1 model exclusively to U.S. entities.

The request comes despite AISI’s established status as a leading global authority in frontier model safety evaluation. The U.K. institute previously evaluated OpenAI’s GPT-6 Astra prior to release and recently published findings on autonomous agent deception. Meanwhile, the U.S. equivalent agency, the Center for AI Standards and Innovation (CAISI), faces operational capacity constraints, operating without a permanent director and maintaining a small staff.

Diplomatic responses to the policy shift remain cautious, with U.K. Prime Minister Andy Burnham advocating for unified international evaluation standards during remarks at the UN General Assembly. However, the directive signals a growing push by Washington to assert national control over early access to advanced artificial intelligence capabilities.

Why it matters

  • Frontier AI labs face tightening domestic regulatory constraints that prioritize national security reviews over international safety collaboration.

  • Geopolitical divergence in model access timelines could delay global deployment schedules and complicate cross-border compliance.

  • Resource shortages at US reviewing bodies like CAISI may create bottlenecks for model release schedules.

Source: the-decoder.com