AISI access to frontier models is back: Reuters reported on October 11, 2026, citing the Telegraph, that the Trump administration has reversed a decision that had asked developers such as OpenAI and Anthropic not to share new models with Britain's AI Security Institute before release. In plain terms, British evaluators can once again test unreleased American models for capabilities and misuse risks ahead of launch. Spokespeople for the White House, OpenAI and Anthropic did not immediately respond to Reuters requests for comment.
How the freeze happened, and how it ended
The freeze began on September 24, 2026, when Washington asked the labs to hold new models back until it finished its own national security review; the pause lasted about 17 days. The clearest public evidence during that window was Anthropic's Claude Mythos 5.1, which was available only to US organizations while British evaluators were kept out. After lobbying by AI minister Kanishka Narayan and other ministers, Washington changed course, though the October 11 reversal came with no published findings from the security review.
Note the wording: what Washington withdrew was a request, not a law. The pre-release arrangement had rested on voluntary cooperation from the labs, and it was that very voluntariness that made it so easy to pause.
Why pre-release testing carries weight
The AISI was set up after the first international AI safety summit in 2023, and its job is to test the most advanced models, evaluating their capabilities and the risk of misuse. Testing before release is fundamentally different from testing afterward: if evaluators find dangerous capabilities before launch, vendors can still adjust safeguards, restrict access tiers or delay the rollout. Once a model ships, every fix runs against the clock of user replication.
That is why the impact of these 17 days outlasted the pause itself. It sent a signal to every country relying on the same voluntary arrangement: your testing window can be frozen at any moment by someone else's national security caveat, and it can be unfrozen just as quickly, on the strength of diplomatic lobbying rather than due process.
Restored, but no stronger than before
Washington's reversal does look like a softer stance on international safety cooperation, and evaluators being back at work is genuinely good news. But the structural weaknesses remain untouched: this is still not an intergovernmental agreement, there is no binding timetable, and the public does not know what the 17-day review concluded, or whether a similar review could quietly restart in the future.
For countries that depend on imported frontier AI, this episode is a case study. Evaluation rights are themselves a form of sovereignty: if you cannot test a system, you cannot explain to your citizens why they should trust it. The real test of whether AISI emerges stronger will not be how many announcements accompany the reversal, but whether British evaluators are seated at the table when the next model launches.
