UK Worries Anthropic Is Locking It Out Of AI Testing

The global AI safety debate has moved from polite cooperation into a more uncomfortable question: who gets to inspect the most powerful models before the public does? The UK is reportedly worried that Anthropic has limited access to one of its newest systems, leaving Britain’s AI Security Institute outside a pre-release testing process it expected to be part of.

According to the Financial Times, Anthropic made its latest model available to vetted US organisations but did not submit it to the UK’s AI Security Institute before launch. The AISI says its mission is to help governments understand the risks posed by advanced AI, including risks to national security and public safety.

That may sound procedural, but it is politically significant. The UK has tried to position itself as a serious global AI safety hub since the 2023 AI Safety Summit. If leading US labs begin limiting their most sensitive model evaluations to American institutions, then international AI governance becomes less global and more national.

Anthropic has built much of its brand on safety, caution and responsible scaling. That makes the reported restriction especially sensitive. The company can argue that advanced models carry security risks and that access needs to be tightly controlled. But governments outside the US will ask why they should trust models they cannot independently test before release.

This is not only about Britain. It is about whether frontier AI becomes another strategic technology controlled through access rules, export controls and national-security filters. Chips have already moved in that direction. Models may now be following. If so, countries that do not host frontier labs may end up consuming systems they have limited power to evaluate.

The issue connects with recent warnings about AI agents and cyber risk. We recently wrote about why OpenAI and other major players are warning the world to prepare for AI hacks. If the risks are real enough to restrict model access, they are also serious enough to require trusted evaluation across more than one country.

For Africa, the lesson is direct. If the UK, with its technical institutions and diplomatic weight, can be locked out of frontier evaluation, African regulators will need to think hard about their own capacity. AI sovereignty is not only about building local models. It is also about having the expertise to test imported systems before they shape education, finance, health and public services.

The path forward should not be open access to every dangerous capability. But it also cannot be a world where only a few US-approved bodies know what the most powerful models can do. AI safety needs trust, and trust needs more than private assurances from the companies building the systems.