Trump Administration Lifts Restrictions on US Companies Sharing AI Models with UK Security Institute
The Trump administration reportedly lifted restrictions on US AI companies sharing their latest models with the UK AI Security Institute for security assessments, allowing US developers to provide pre-release access to the UK without violating government requirements.
Previously, on September 24, the White House, through the Office of the National Cyber Director, requested OpenAI and Anthropic to refrain from sharing new models with the UK AISI until the US government completed its review, leading Anthropic to limit Claude Mythos 5.1 to US organizations only.
This reversal occurred on October 11, after UK AI Minister Kanishka Narayan lobbied for AISI to regain pre-release testing access to cutting-edge models from Anthropic, OpenAI, and Google DeepMind, which had been suspended for about 17 days.
AISI was established after the first global AI safety summit in 2023, specifically to test the capabilities and abuse risks of advanced models, previously enjoying privileged early access. This restriction had impacted the UK's efforts to lead in international AI regulation.
Trump had expressed at the UN General Assembly a desire to encourage rather than constrain AI, opposing interventions by "globalist" organizations in the development of US company models, while his newly formed AI task force called for developers to report security incidents.
The incident was driven by a balance between US priority reviews and ally testing permissions, with funding and resources flowing towards domestic security reviews and the restoration of UK AISI testing capabilities, benefiting UK institutions seeking early security assessments and willing US companies, while putting pressure on frontier labs that were previously required to prioritize US reviews.
Anthropic had publicly coordinated with the US government to expand access, while OpenAI had previously provided partial pre-release access to AISI; this restoration now covers major frontier developers.
Source: Public Information
ABAB AI Insight
The Trump administration signed an executive order in June 2026 establishing a voluntary framework that allows cutting-edge models to submit for federal government review of cybersecurity capabilities up to 30 days before release, evaluated through agencies like CAISI. Historical trends show a shift from opposing strict security requirements to limited domestic priority reviews, while maintaining voluntary cooperation with companies like OpenAI and Anthropic.
On the capital front, the US views access to cutting-edge models as a national security asset, prioritizing domestic reviews, with resources shifting from early multilateral sharing to a sequential approach (US before allies). The motivation is to ensure the security of US systems before sharing, while the reversal avoids overly damaging cooperation with the UK and maintains laboratory operational flexibility.
Similar cases include voluntary security commitments during the Biden administration and early access arrangements with the UK AISI following the Bletchley summit. The US is currently in a transitional phase from domestic priority control to limited international cooperation recovery, while the UK is attempting to maintain its testing leadership through lobbying.
Essentially, this reflects regulatory changes and a shift in pricing power: pre-release model access has transitioned from shared resources among allies to a sortable security asset for the US. The mechanism involves administrative directives and lobbying dynamics making access rights a reversible policy tool, reinforcing US priority insights into frontier capabilities while maintaining a superficial transatlantic cooperation through short-term reversals, avoiding a complete severance of international security assessment channels.
ABAB News · Cognitive Laws
Models are first shown to themselves, then to allies.
Restrictions are directives, while lifting is lobbying.
Security assessment rights are tradable priorities.