White House restores UK AI institute’s access to frontier models

The ban and the reversal

The White House’s September order cut off the institute’s access to models from companies that had been providing them. Anthropic’s restriction on Claude Mythos 5.1 made the consequences concrete; the model was new, and the institute could no longer examine it.

Narayan plans to announce new AI safety measures in an upcoming speech. They are expected to include stricter reporting requirements for AI incidents affecting critical infrastructure.

Two views of who should set the rules

The access dispute sits alongside a broader disagreement about international oversight. UK Prime Minister Andy Burnham previously backed international AI regulation at the UN General Assembly. At the same event, Donald Trump warned that “globalist” organizations should not interfere in the development of American AI.

Leaders of MI6, MI5 and GCHQ have also warned ministers that AI is expanding the capabilities of hostile forces and terrorists.

I think the reversal matters less as a settled policy than as evidence of how fragile access for safety research can be. The announcement leaves open what would prevent another cutoff, and whether the planned reporting rules will apply to incidents involving systems that independent evaluators cannot inspect.