The White House on Oct. 11 reversed a Sept. 24 Office of the National Cyber Director directive that had instructed Anthropic, OpenAI, and Google DeepMind to withhold frontier models from Britain’s AI Security Institute pending a US national security review. The freeze lasted 17 days. Anthropic complied with the original order, blocking UK officials from reviewing Claude Mythos 5.1 for the duration.

UK AI Minister Kanishka Narayan confirmed Monday that AISI is “now conducting rigorous evaluations of the latest generation of models, including Gemini 4 Argon.” The Decoder, citing the Daily Telegraph, attributed the reversal to diplomatic pressure from Narayan himself.

Read what Narayan said about the episode: “When you’re reliant on someone else’s technology, you have to be clear-eyed about the permanence of any outcome.” That’s a minister of a G7 government describing dependency on three American companies the way a founder might describe a single-source supplier.

No statute governs this access arrangement. One ONCD memo closed it; another opened it. The next could close it again. For a 1-to-4-person business building customer acquisition on top of Claude or Gemini, that’s the point: the model layer is a political instrument, and vendor terms don’t bind Washington.

This sits downstream of mandatory AI incident reporting and runs parallel to the administration’s AI liability posture. Narayan is expected to announce stricter AI incident reporting rules for UK critical infrastructure next.

LemonLime prepares finished customer-growth work, leads, outreach, content, as a done-for-you service, so an owner isn’t the one reconfiguring workflows when model access shifts.

The reversal doesn’t retire the risk. It documents it.

Sources

Sources