Anthropic withheld Mythos 5.1 from the UK's AI safety testers

Anthropic did not provide Claude Mythos 5.1 to the UK's AI Security Institute for pre-release testing, according to the Financial Times.

The decision is notable because Anthropic has previously worked closely with the UK institute on evaluations of advanced AI systems.

Mythos 5.1 is Anthropic's most capable model for specialized cybersecurity and biology research.

Access is currently restricted to a small number of vetted US organizations.

Why Mythos 5.1 is treated differently

Anthropic says Mythos models require tighter controls because of their capabilities in areas that can be useful for both defensive research and harmful activity.

The model is designed for advanced cybersecurity and life-sciences work.

Anthropic's own documentation says Mythos 5.1 currently remains available only to selected US organizations while the company works to expand access.

The UK AI Security Institute, commonly known as AISI, has previously evaluated frontier models before release.

The Financial Times reports that UK officials are concerned that losing access to systems at the frontier of AI research could make it harder for the country to independently evaluate their risks.

Frontier AI access is becoming geopolitical

The issue extends beyond one model.

The world's most capable AI systems are increasingly being treated less like ordinary software products and more like strategically sensitive technology.

Governments are introducing restrictions around advanced chips, cybersecurity tools and other technologies that could affect national security.

Frontier models may increasingly fall into the same category.

Anthropic said it is coordinating with the US government while working toward broader access.

There is no evidence that Anthropic has permanently excluded UK researchers from Mythos.

But the episode demonstrates a new tension in AI safety.

Governments want independent access to powerful models so they can understand the risks.

The companies building those models increasingly have reasons to restrict who gets access in the first place.