The UK AI Security Institute (AISI) has regained access to pre-release frontier AI models from US developers, concluding a 17-day period during which access was frozen. The interruption, which began around September 24, 2026, was reportedly due to a request from the Office of the National Cyber Director prioritizing US safety reviews.
Access was restored on October 11, 2026. During the freeze, Anthropic notably withheld its Claude Mythos 5.1 model from non-US entities. UK AI Minister Kanishka Narayan is credited with playing a role in securing the renewed access for the AISI.
Background on AISI and Voluntary Agreements
The AISI, established in 2023 following the international AI Safety Summit at Bletchley Park, operates under voluntary agreements. These agreements, struck in 2023, mean the institute’s access to cutting-edge AI systems depends on developers choosing to share their models. The institute employs over 100 technical experts focused on safety evaluations. Previously, the AISI evaluated OpenAI’s GPT-6 Astra, an effort that contributed to its global role in assessing frontier models.
National Priorities vs. International Collaboration
The recent interruption highlighted the tension between international safety testing commitments and national priorities. When the Office of the National Cyber Director requested that domestic safety reviews take precedence, the existing framework adjusted accordingly. Anthropic’s decision to restrict access to its Claude Mythos 5.1 model for non-US entities during this period illustrated this dynamic, demonstrating how companies may balance global safety commitments with signals from their home government.
The restoration of access may alleviate concerns that AI safety testing could become fragmented along national lines. According to the article, this development could reassure global researchers and technology investors who may have worried about a divergence between US and UK approaches to AI governance. The episode is also seen as setting a precedent that other nations pursuing international AI regulation through collaboration may observe.
Future Outlook and Remaining Uncertainties
Among the models now under review by the AISI is Google’s Gemini 4 Argon. Whether Anthropic’s Claude Mythos 5.1 will subsequently be shared with non-US entities is considered an early indicator of how normalized international access has become.
Looking ahead, several uncertainties remain. It is unclear whether Anthropic will share its Claude Mythos 5.1 model with non-US entities. Additionally, questions persist about whether the US will formalize priority rules for domestic safety reviews, if more developers will resume sharing models with non-US evaluators, and whether the current voluntary framework from 2023 will be replaced by a more robust structure. The source states that AISI officials have pointed to future collaborations focused on using advanced AI to strengthen cyber defenses, an area that aligns with the concerns of the office that initiated the review prioritization.
The timeline of events includes the establishment of AISI and its voluntary agreements in 2023, the approximate start of the access freeze on September 24, 2026, and the restoration of access on October 11, 2026.
Why This Matters
The materials describe a narrow update: The UK AI Security Institute had its access to pre-release frontier AI models from US developers frozen for 17 days, starting around September 24, 2026, due to a request from the Office of the National Cyber Director prioritizing US safety reviews. Whether Anthropic’s Claude Mythos 5.1 will be shared with non-US entities.
Broader Context
Source materials place the factual news in this context: AISI was set up in 2023, following the international AI Safety Summit held at Bletchley Park.
