The freeze, reconstructed

The order was never published, so the paper trail is patchy by design. What we know: Politico reported on September 24 that the Office of the National Cyber Director asked OpenAI and Anthropic to hold new models from the UK's AI Security Institute pending a US review. A British official confirmed the request to Bloomberg the next day. The reviews were to run through CAISI, the US testing body, which at the time had no permanent director and a few dozen technical staff, according to Tech Times. A gatekeeper with barely a gate.

Only one model can be named with confidence. Anthropic restricted Mythos 5.1 at launch to US organizations, and its own release materials say so outright. Everything else is inference. The AISI letter to Parliament, dated a week before the freeze, confirms the UK had pre-release access to GPT-6 Astra. So the request either landed after that model was already out the door, or the labs read it differently, or OpenAI simply moved faster than the paperwork. None of those readings is flattering to the process.

Who blinked

The Telegraph says lobbying from AI minister Kanishka Narayan and other ministers did it. Reuters frames the reversal as a softer stance on international AI safety cooperation. There is also a Friday data point worth keeping next to this one: the new AI task force said developers must report security incidents with their models, which [we covered separately](https://alfw.ai/article/white-house-mandatory-ai-incident-reporting/). Put the two together and the pattern is legible: tighter leash at home, the tap reopened abroad. The through-line in both cases is that Washington wants to see things first.

The part that survives the reversal

Here is the durable lesson, and it has nothing to do with Britain. The UK's testing access rests on negotiated agreements with the labs, not statute. Brussels, by contrast, is building binding obligations: models designated as systemic risk face mandatory assessments, with fines behind them. We covered [the EU's safety questionnaires to thirty labs](https://alfw.ai/article/eu-30-ai-labs-safety-questionnaires/) earlier today. One system runs on law. The other ran on an understanding, and that understanding got suspended for 17 days by a single office in Washington.

London got its access back. The lesson about what informal deals are worth survives either way.

For the labs, the takeaway is simpler. Anthropic did what Washington asked and paid for it in London's trust. The next time a government asks for something quietly, the labs will remember that the quiet part got loud.

Sources

  1. [1] Startup Fortune, “White House Restores UK Access to Frontier AI Models After 17-Day Freeze” (October 11, 2026)Read source
  2. [2] Reuters wire, “US reverses course on Britain's testing of AI models, Telegraph says” (October 11, 2026)Read source
  3. [3] AI Tool Bible, Mythos 5.1 and AISI letter backgroundRead source
  4. [4] Tech Times, CAISI staffing (September 28, 2026)Read source
  5. [5] Forkast analysis (September 25, 2026)Read source