What is Narrow AI (Weak AI)?

The White House has asked OpenAI and Anthropic to withhold their latest AI models from Britain’s safety watchdog until they have been tested by the US government, threatening the UK’s prized early access to frontier tech.

The request from the US Office of the National Cyber Director means the American companies have been asked to give Washington first look at new models before sharing them with the UK’s AI Security Institute (AISI).

A senior US administration official said Washington wanted to ensure American systems were secure before models were shared with allies.

“Because they’re American companies and this has been our policy with every new frontier model that comes out”, the official said.

Anthropic appears to have already followed the request. The company did not provide its latest Claude Mythos 5.1 model to AISI, saying access was initially limited to US organisations.

“We’re coordinating with the US government to expand access to a broader set of domestic and international partners as quickly as possible”, Anthropic said.

AISI has built close relationships with leading developers to test their most powerful models before release, searching for vulnerabilities including their ability to conduct sophisticated cyber attacks or assist with biological threats.

AISI director Henry de Zoete acknowledged this month that the watchdog had not received Anthropic’s latest model, but said it retained “strong relationships with all frontier AI developers”.The institute tested OpenAI’s GPT-6 Astra before its release, he added.

Pressure mounts on AI watchdog

The restrictions come as AISI faces growing pressure over the pace of advances in AI.

Multiple staff have reportedly been signed off work with stress and received counselling, with current and former employees pointing to demanding testing schedules and alarm over increasingly capable systems.

AISI chief technology officer Jade Leung is also stepping back from her full-time role at the end of September due to “personal circumstances”, although she will remain as vice-chair and become security adviser to the government’s AI Taskforce.

Leung has previously warned that governments risk falling behind the technology, while AISI researchers have found vulnerabilities in every advanced system they have tested.

The US move also comes amid growing calls from AI executives themselves for international cooperation. OpenAI boss Sam Altman and Anthropic chief Dario Amodei both called for common global safety standards at the UN this week.

But the Trump administration has rejected new global AI governance, highlighting a growing divide over how the world’s most powerful models should be controlled.

The intervention could complicate Andy Burnham’s push to make AI safety a focus of Britain’s G20 presidency next year, where he has called for a “single set of global principles and standards” for the technology.