White House restores UK AISI's access to models from Anthropic, Google and OpenAI

The White House has reversed its own decision and restored the UK's AI Security Institute's early access to frontier models from three labs: Anthropic, Google DeepMind and OpenAI. In September, the Trump administration told the companies to stop sharing models with the British agency. The reversal came within weeks and, according to The Decoder, was won through diplomatic pressure from UK AI minister Kanishka Narayan.

What happened

The publicly available facts are thin, but the picture is clear. Washington first shut the channel that gave the British institute advance access to new models, then opened it again. Formally, the directive targeted the companies, not the institute itself: Anthropic, Google DeepMind and OpenAI were the ones told to halt the transfers. The restriction is now lifted, and the labs can resume their previous practice.

The original headline calls the body the AI Safety Institute, while the article text calls it the AI Security Institute. The second name is current: the British agency was renamed in February 2025, shifting its focus from general "safety" to threats to national and public security.

Why the institute needs early access

The point of such access is that a model can be tested before public release. Government testers run it through risky scenarios: help with cyberattacks, biological threats, the ability to bypass safeguards. The results go back to the labs while fixes are still cheap to make. Without pre-release access, the institute sees a model only after millions of people are already using it.

For the British side, this is a question of the institute's weight. If the best models reach it last, or not at all, its assessments lose value and it becomes a mere commentator. For American labs, the calculation is different: they are better off with one clear testing regime than with separate arrangements with each government.

What is known and what is not

Key details from the report:

• Who was affected: the UK's AI Security Institute.
• Who the directive applied to: Anthropic, Google DeepMind, OpenAI.
• When the ban was introduced: September 2026.
• What was done now: early access was fully restored.
• Who secured the reversal: UK AI minister Kanishka Narayan.

It is not known how the White House justified the September ban. It has not been disclosed whether new conditions were set, for example which models fall under the access arrangement, whether test results must be cleared with US agencies, and whether the ban will survive in the form of exceptions. There are also no public comments yet from the labs themselves. Without these details, it is too early to consider the matter closed: the decision could be reversed again as quickly as it was made.

What teams and regulators should do

Companies that release or deploy frontier models should build a buffer for external testing into their release plans and not assume this channel is stable. The political weather between Washington and London changes faster than model development cycles.

Enterprise customers should ask vendors whether a specific model went through pre-release testing by government institutes, and which ones. It is one of the few independent signals of how mature a model's safeguards are.

For regulators in other countries, the episode shows what access to the best models rests on: informal agreements and political will, not law. Those building their own AI evaluation institutes will have to put access rules in writing, or a single order from Washington could wipe out all their work again.