White House requests delay AISI, Anthropic restricts access

The White House has asked OpenAI and Anthropic to share new AI models only after evaluation by U.S. agencies with the AI Safety Institute (AISI) in the United Kingdom. This request stems from the Office of the National Cyber Director and concerns the tests AISI conducts with highly advanced systems, known as frontier models, which allow early access for regulators. Anthropic has since decided to restrict Claude Mythos 5.1 to U.S. organizations only.

This step places both OpenAI and Anthropic in a difficult position: either hold back the models for AISI or risk unsettling the White House. Anthropic has chosen the option to limit the model to the U.S., with the Center for AI Standards and Innovation (CAISI) serving as the primary U.S. point of contact. However, this organization faces shortcomings as it lacks a permanent director and has only a few dozen employees. This situation may lead to delays in the necessary model evaluations.

AISI functions as a well-equipped agency for AI testing and has early access to leading frontier models. Since its establishment, AISI has made significant findings, including reports of autonomous deception by AI agents in real-world environments. In a recent letter to the British Parliament, Director Henry de Zoete indicated that AISI still has access to the frontier models and that they have already tested OpenAI’s GPT-6 Astra prior to its release. Prime Minister Andy Burnham sought to temper tensions during the UN General Assembly by calling for a joint approach with global principles and standards.

Read the full article from AI Insider.