Britain’s AI safety lab found a U.S. border around Anthropic’s pre-release access

Anthropic declined to give the UK AI Security Institute pre-release access to Claude Mythos 5.1, Ross Kelly at IT Pro writes, citing the Financial Times. The outlets report this is the first time AISI has been left out of an Anthropic pre-release evaluation.

Anthropic launched Mythos 5.1 on September 1 for vetted users. It uses the same underlying model as generally available Claude Fable 5.1, with more permissive safeguards for cybersecurity and life-sciences work. Project Glasswing is Anthropic's invitation-only defensive-cyber initiative; life-sciences access runs through a separate verification program.

AISI is the UK's government-backed lab that tests powerful AI models around release. Pre-release testing means outside evaluators poke the model before the wider public gets it. The institute had access to Mythos Preview in April.

Similar U.S. organizations still got in, per the FT reporting through IT Pro. UK officials fear a wider protectionist shift among U.S. tech companies.

Anthropic had not explained the decision at publication. IT Pro approached the company and got no response.

A UK Cabinet Office spokesperson told IT Pro that the UK remains a world-leader in AI security, that AISI continues to collaborate closely with industry partners including Anthropic, and that only last week AISI tested OpenAI's most powerful model, GPT-6 Astra, before public release. Risks, the spokesperson said, do not stop at national borders.

June brought temporary U.S. export restrictions on earlier Mythos and Fable models for foreign users, later eased after safeguard work. Neither source identifies those earlier restrictions as the reason AISI was excluded.

Pre-release testing depends on developers granting access. The reported exclusion shows how a U.S.-only rollout can limit a foreign government tester's reach.

AISI evaluated OpenAI's GPT-6 Astra before release. Mythos 5.1 remained limited to a set of U.S. organizations.

The safety test stops where access stops.

Sources

• IT Pro — Anthropic withholds Mythos 5.1 from UK AISI
• Financial Times — scoop (via IT Pro)
• Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1
• Anthropic — Project Glasswing
• Anthropic — Redeploying Claude Fable 5
• AISI — Our evaluation of Claude Mythos Preview’s cyber capabilities
• OpenAI — Path to Astra