Britain’s AI safety testers are back in the room. The UK’s AI Security Institute (AISI) can once again evaluate advanced models from major US developers, including OpenAI and Anthropic, before those models reach the public.
The shift follows a reversal by the Trump administration on October 11, 2026. It restores a voluntary arrangement that Washington had effectively paused only weeks earlier.
What changed, and why it matters
The story starts in September 2026. That month, the White House asked OpenAI and Anthropic to hold back new models from AISI until US reviews had been completed.
The result was a sequencing rule. American officials got first look, and British evaluators waited in line until Washington signed off.
The directive had real consequences. Anthropic restricted the release of its Claude Mythos 5.1 model to US organizations, based on the White House request.
The October 11 reversal unwinds that. US developers can again share pre-release models with AISI under the voluntary access setup that existed before the September directive.
That means British testers can probe models for safety risks and potential misuse ahead of launch, rather than after the fact.
The lobbying behind the reversal
The change did not happen on its own. UK AI Minister Kanishka Narayan pushed for British evaluators to regain access.
Narayan’s case leaned on the UK’s existing working relationships with developers like OpenAI. The argument, in effect, was that those ties were built on trust and early access, and cutting them off undercut both.
The pre-release access regime is voluntary, which means it depends on the goodwill of developers and, as September showed, the preferences of the White House.
Background: Bletchley Park and the birth of AISI
AISI traces its origins to the AI Safety Summit held at Bletchley Park in November 2023. The institute’s core job is evaluating advanced AI systems, with a particular focus on identifying potential misuse risks associated with capabilities related to cybersecurity and biological threats. Its findings feed into UK policy on AI governance.
The UK is also weighing proposals for mandatory incident reporting. That would add a compulsory layer on top of the current voluntary testing arrangements, though the proposals remain under consideration rather than adopted.
What this means
For the developers, the reversal removes a potential headache. Anthropic’s decision to limit Claude Mythos 5.1 to US organizations showed how a government request can ripple into product availability, and a return to the earlier arrangement reduces the odds of fragmented, country-by-country launches.
Washington’s focus has been domestic security, ensuring that US officials review sensitive capabilities first. London’s focus has been cementing its status as an independent, trusted evaluator that companies and governments can rely on.
Because the arrangement is voluntary, a future policy shift in Washington could restore the withholding requirement. Mandatory incident reporting, if adopted, would give British regulators a channel that does not depend entirely on developer cooperation or White House approval.
Leave a comment