Five days after our coverage of the White House AI safety accord, the U.S. government reversed an earlier request and will allow American AI firms to give Britain's AI Safety Institute early access to frontier models for safety testing.
Until this week, U.S. officials had asked companies not to share unreleased models with the UK institute. Reuters, citing the Telegraph, reports the administration has changed that stance, so developers like OpenAI and Anthropic can now grant early access without running afoul of U.S. policy.
That matters because who tests advanced models determines how quickly dangerous capabilities and misuse risks are discovered. Britain set up the AI Safety Institute after the 2023 global summit to probe the most capable systems, and early, controlled access lets independent experts examine models before they reach broad users.
Think of it like a manufacturer letting a trusted inspector look at a prototype behind closed doors rather than waiting for public launch. The change does not make models public; it simply clears a path for controlled, cross-border safety checks.
Practically, U.S. labs can cooperate with British testers sooner, which should speed discovery of hard-to-find safety flaws. It also signals governments are carving out ways to share evidence about high-risk systems across borders.
The open question is whether this limited cooperation will translate into faster fixes and shared standards, or whether practical hurdles and trust limits will keep most testing inside national lanes. Watch how quickly firms actually use this permission and what findings the UK institute publishes.
