Open-weight AI models closing gap to frontier, but raising safety risks

The UK's AI Safety Institute released an analysis showing that open-weight AI models—those whose code is publicly released—are narrowing the performance gap with frontier closed commercial models. The institute found that the most capable open models currently lag four to seven months behind the best closed models when measured on cyber capabilities. Open-weight models offer concrete benefits: they can be adapted and fine-tuned for specific organizational or user needs without reliance on a third-party provider, and they facilitate easier collaboration across teams and institutions. However, this openness comes with a critical liability. Because the models are publicly available, users can remove or disable the safety guardrails and restrictions that creators have built in, making advanced AI capabilities—particularly in cyberattacks and biological threat research—available to actors with harmful intent. Evidence of this risk is already materializing: The New York Times reported that Boko Haram militants in Nigeria are leveraging AI tools (predominantly closed models, in that case) to plan military attacks and gain tactical advantages. Anthropic CEO Dario Amodei recently acknowledged in writing that open models "potentially present a higher risk," yet he framed the safety challenge as broader than the open versus closed distinction, arguing that all powerful AI models, regardless of their licensing model, should be evaluated and tested for safety risks before deployment.