White House pressures OpenAI and Anthropic to give US agencies first look at new AI models before UK's AI Safety Institute

The White House's Office of the National Cyber Director asked OpenAI and Anthropic to let US agencies review new AI models before sharing them with the UK's AI Safety Institute (AISI). Anthropic complied, restricting Claude Mythos 5.1 to US organisations only. OpenAI appears not to have fully complied: AISI says it still got early access to test GPT-6 Astra before release. No dates, formal scope or enforcement mechanism for the request have been disclosed.

This is a real shift in who gets pre-deployment access to frontier models, not just rhetoric: one major lab has already changed its release behaviour. If it holds, UK regulators, enterprises and researchers relying on AISI's independent testing lose a source of early, non-US safety evaluation, and other labs may face similar pressure to prioritise US review. The catch is that the US body meant to absorb this role, CAISI, is reportedly under-resourced, so the practical safety benefit of the request is unclear even if the political effect is real.

The "geopolitical rift undermining international AI safety cooperation" framing is interpretive, not established fact. The evidence chain is thin: a Decoder story built on Politico reporting, with only a bare domain link as sourcing, no named documents, and no stated scope, duration or legal basis for the request. Calling this a clear breakdown in cooperation overstates what two data points (one compliant lab, one apparently non-compliant lab) actually show.

1. 01Whether OpenAI's next major model release also gets AISI early access, or whether GPT-6 Astra was a one-off before compliance tightens
2. 02Any formal statement, executive order, or ONCD memo defining the scope and duration of the US-first review request
3. 03Whether AISI or the UK government responds publicly (e.g. a statement from Andy Burnham or a Parliamentary reply referencing Henry de Zoete's letter) confirming reduced access