What happened
The White House Office of the National Cyber Director (ONCD) formally directed OpenAI and Anthropic to withhold their newest frontier AI models from the UK's AI Security Institute (AISI) until the US government completes its own domestic review, first reported by Politico around 24 September 2026 and confirmed by the UK to Bloomberg. Anthropic has already complied — its Claude Mythos 5.1 (released 1 September 2026) was made available only to US-vetted organizations via Project Glasswing, marking the first break in years of pre-release AISI evaluations. Reviews are now routed through CAISI, which has had no permanent director for ~67 days and minimal standing evaluation infrastructure.
Why it matters
This treats pre-release access to the most powerful AI systems as a sequenced US national-security asset rather than a shared allied resource, formalising the template set by June 2026 Commerce export-control orders. It weakens independent, cross-validated safety evaluation (US CAISI + UK AISI) at the exact moment enterprises and governments rely on published pre-release evaluations to inform procurement of frontier models.
Action needed
AI deployers/government buyers should not assume multi-body independent evaluation of new frontier releases; factor reduced independent evaluation coverage and US-first sequencing into procurement and risk decisions for frontier model adoption.