Sources: the US restores the UK AISI's access to the latest US models for safety testing, after ordering OpenAI and Anthropic to stop sharing models last month

US advanced AI models are now available for UK safety testing, which may accelerate the discovery of vulnerabilities before widespread deployment.

The UK's AI Safety Institute (AISI) has regained access to the latest US AI models for safety testing, a privilege temporarily revoked last month when the US ordered OpenAI and Anthropic to cease sharing their models. This restoration comes as UK AI Minister Kanishka Narayan is set to emphasize the need for Britain to develop its own AI infrastructure rather than solely relying on external tech companies. The move follows concerns raised by Geoffrey Irving, a former chief scientist at AISI, who warned about the potential for adversarial states to steal cutting-edge AI models due to insufficient security measures at AI firms. Irving also highlighted the rapid progress of AI, suggesting a "50% chance" of superintelligence emerging within a year and advocating for a pause in advanced AI development.

The UK's AI Safety Institute (AISI) regaining access to advanced US AI models signals a pragmatic approach to AI regulation, prioritizing safety testing over immediate concerns about model proliferation. This suggests a potential divergence from a complete lockdown strategy, acknowledging that international collaboration, even with access restrictions, is necessary for robust AI safety evaluations. The situation highlights the delicate balance governments are attempting to strike between fostering innovation and mitigating existential risks, with the UK taking a cautiously engaged stance.

Minister Narayan's call for increased domestic AI infrastructure and the previous security warnings from AISI’s former chief scientist underscore a growing recognition of the strategic importance of AI sovereignty. The incident where a US company's AI model attempted to assist with a murder investigation further amplifies anxieties about AI's autonomous capabilities. These events collectively point towards an intensified focus on AI governance, with an expectation of greater scrutiny on model behavior and a push for more independent safety verification mechanisms globally.