On September 25, Politico reported that the US White House has mandated that OpenAI and Anthropic must first undergo security evaluations by the US government before allowing the UK's Artificial Intelligence Safety Institute (AISI) to test their latest AI models. This stipulation, put forward by the US Office of the National Cyber Director, is designed to ensure that the US government can assess cutting-edge AI models as a priority, thereby protecting US systems before granting access to allies.

Currently, Anthropic has halted the provision of its Claude Mythos 5.1 model to the UK's AISI, stating that it is working in tandem with the US government to progressively broaden access for both domestic and international partners. Previously, the UK's AISI had a history of collaborating closely with AI firms like OpenAI and Anthropic to conduct pre-release security testing on their models.

This demand from the US government comes amid its ongoing exploration of how to effectively regulate rapidly evolving advanced AI models. Earlier, certain AI companies and experts cautioned that insufficiently tested state-of-the-art models could introduce security vulnerabilities. On the contrary, the UK is steadfast in its commitment to fostering the creation of globally harmonized AI safety standards. UK Prime Minister Andy Burnham underscored the necessity for international collaboration in AI safety, rather than a unilateral approach by any single nation.

Meanwhile, the US is bolstering the capabilities of the AI Standards and Innovation Center at the Department of Commerce's National Institute of Standards and Technology, with the aspiration that it will emerge as a pivotal entity in US AI safety governance. OpenAI also expressed this week its intent to collaborate with international bodies, including the UK's AISI, to collectively advance the establishment of global AI safety standards.