US restores UK AI Security Institute's early access to models after 17-day pause

The United States has given Britain's AI Security Institute (AISI) back its early look at unreleased artificial intelligence models from OpenAI, Anthropic and Google DeepMind. The reversal took effect on Sunday, 11 October 2026, ending a pause of 17 days. On 24 September, the White House's Office of the National Cyber Director had asked the three American labs to stop sharing new models with the British testers while Washington completed a national security review of its own. The decision was reported on Monday, 12 October 2026, and is being credited to lobbying by Kanishka Narayan, the UK's minister for AI.

Pre-release testing is one of the few practical checks that governments have on the most powerful AI systems. Before a frontier model reaches the public, government evaluators probe it for risks such as help with cyber attacks or with dangerous chemistry and biology, and the findings can shape the safeguards a company ships. Britain's institute is among the best known of these testing bodies, and its work has been cited in the published safety documents of leading labs. Losing access, even for a few weeks, meant British evaluators could not see the newest systems at the point where their input matters most.

The episode also shows where the real control sits. The arrangement under which labs share models with foreign testers is voluntary, and it rests on the goodwill of the US government, which can ask companies to hold back at any time. That is why some observers have warned that the restored access should not be treated as permanent: a request made once can be made again. For a field that is moving fast, a gap of even a few weeks can mean that a major model is released without outside evaluation.

For India, the lesson is direct. Indian policy researchers have already argued for a trusted India-US corridor that would secure access to frontier AI. The British experience suggests that such access depends on political relationships as much as on technical capacity, and that countries without a formal agreement can be cut out quickly when Washington's security concerns rise.

Short-term: British evaluators can resume testing the newest systems from the three labs, including models they missed during the pause. Companies that rely on AISI findings for their own safety reports regain an outside reviewer.

Long-term: The pause is likely to push the UK and other allies to seek written, binding arrangements rather than goodwill-based ones. It may also speed up efforts by other governments, India among them, to build their own testing capacity instead of depending on access granted from Washington.

Who is affected: Safety researchers at the UK AI Security Institute; OpenAI, Anthropic and Google DeepMind; British and allied policymakers; businesses that use AISI evaluations to judge model risk; and governments, including India's, that are planning their own frontier AI testing.

Questions and Answers