Sources: US restores UK AISI's access to latest US models for safety testing after ordering OpenAI and Anthropic to stop sharing models.
ai-safety anthropic deepmind google openai
| Source: Techmeme | Original article
The United States has reinstated the UK’s AI Safety Institute’s access to the latest American AI models for safety testing, after previously ordering OpenAI and Anthropic to halt model sharing.
The United States has lifted a 17‑day ban on the United Kingdom’s AI Security Institute (AISI) receiving pre‑release versions of frontier AI systems. The pause, imposed in September, barred British officials from examining the latest models from Anthropic, OpenAI and Google DeepMind. As of 11 October 2026, the White House has reversed the directive, restoring AISI’s access for safety testing.
The brief freeze highlighted growing friction over cross‑border oversight of powerful generative models. By preventing regulators from reviewing cutting‑edge systems before they entered the market, the United States risked creating a blind spot in the global safety chain. Restoring access signals a willingness to coordinate on risk assessment and to keep transatlantic standards aligned, even as political pressures in Washington have occasionally pushed for tighter export‑type controls.
The move matters because early‑stage testing can surface alignment, robustness and misuse risks that are harder to address after deployment. It also reassures European partners that the United States will not unilaterally block collaborative safety work, a concern that surfaced after the earlier order forced Anthropic, OpenAI and DeepMind to halt model sharing with Britain.
Going forward, observers will watch whether the United Kingdom negotiates formal data‑sharing agreements or joint audit frameworks with the United States and the AI firms involved. Further policy signals from the White House or the UK government could shape the durability of this cooperation. Additionally, any new restrictions on other allied regulators, or a repeat of the freeze, would test the resilience of the emerging transatlantic AI safety architecture.
Sources
Back to AIPULSEN