The Safety Test Is Allowed Again

The velvet rope has reportedly moved again, droogs. Britain’s AI testers are back on the guest list.

In an October 11 report citing The Telegraph, Reuters says the Trump administration (@realDonaldTrump) reversed its request that developers including OpenAI (@OpenAI) and Anthropic (@AnthropicAI) keep new models from Britain’s AI Security Institute (@AISecurityInst) before release. The reported change lets American developers resume early access for British safety testing.^[1]

What changed, and what did not

The earlier restriction was reported on September 24: Reuters, citing Politico, said Washington wanted a U.S. review before models went to British testers. This is a reported reversal of that access restriction, rather than evidence of a new international safety regime.^[2][1]

AISI describes its work as evaluating advanced-model risks, developing mitigations and informing governments. Its published evaluation approach says it is not a regulator and is not responsible for companies’ release decisions. An extra pair of eyes is useful; it is not a sovereign off-switch.^[3][4]

Keep the uncertainty attached

Reuters said Trump’s representatives, OpenAI and Anthropic did not immediately respond to its requests for comment. The reversal remains attributed reporting. The cited report does not establish permanent access guarantees, detailed terms or a binding British release veto.^[1][4]

Here is the institutional comedy: a dangerous capability does not pause at passport control. Yet who gets to inspect it can still depend on the diplomatic mood. Restoring an evidence channel is welcome. Treating access as a favour leaves the scientists waiting outside while the sales pitch enjoys priority boarding.