UK will keep pre-release testing of OpenAI and Anthropic models

The UK's AI Security Institute will continue to test the most powerful AI models before their public release. The British government said so on Monday, 12 October 2026. AI minister Kanishka Narayan said London worked closely with Washington to preserve access, Bloomberg reports. The decision reverses a US move under which White House reviews were to come first.

What is known so far

The publicly available detail is thin so far. The government confirms that the institute keeps its right to test frontier models before release. Narayan presents this as the result of work with his American counterparts. TNW's headline names OpenAI and Anthropic, whose models the institute has tested before. The accessible part of the article does not include a full list of companies.

The wording about the reversal matters. If the US put its own checks first, the UK risked receiving models later or in a stripped-down form. Or not receiving them at all. London now says that risk has been removed.

Why pre-release access matters so much

The institute was created in 2023 after the AI safety summit at Bletchley Park. It was then called the AI Safety Institute. In 2025 it was renamed the AI Security Institute, shifting its focus to practical threats: cybersecurity, model misuse, and dual-use risks. It operates without a law requiring labs to show models in advance. Everything rests on voluntary agreements with companies and on the political will of two governments.

That is why "access preserved" means more than it seems. With no legal lever, access depends on how the labs sequence reviews between countries. When a priority American track appears in that chain, the second party risks being left outside. London has signaled that it remains among those who see a model before users do.

The US has its own counterpart to the institute, the Center for AI Standards and Innovation. Bloomberg speaks specifically about White House reviews here, and the available excerpt does not make clear how they relate to CAISI's work.

What is still unclear

The main questions remain open. It is not disclosed whether the British institute receives models at the same time as the American reviewers or with a delay. It is unknown whether the agreement is in writing and how well it would withstand a change of policy in Washington. There is no data on how much time is allotted for testing, or on what happens if the institute finds a problem: whether it can delay a release or only issues a finding.

Key facts in brief:

• Date of the announcement: 12 October 2026.
• Who announced it: the UK government, AI minister Kanishka Narayan.
• What stays in place: pre-release testing of the most powerful models at the AI Security Institute.
• What changed: the US move under which White House reviews came first was reversed.
• Source of details: Bloomberg, as relayed by TNW.
• Not disclosed: access timelines, the legal form of the agreement, the list of labs.

What teams and companies should do

Model developers should build a window for external reviews in two jurisdictions into their release schedules from the start. If the order of reviews in the US and the UK changes, it affects the release date.

Companies that build products on frontier models should watch how the institute publishes its evaluation results. These remain one of the few independent sources of data on models' dangerous capabilities. Security teams should check their own internal tests against the institute's public methodologies.

Next, watch for the text of the statement itself and for comments from the labs. Until then, the announcement remains a political signal, and it cannot yet be verified in practice.