An off switch for dual use knowledge in AI models

Anthropic researchers have developed a method to systematically isolate and deactivate specific dangerous capabilities within large language models without degrading their general performance. By identifying the internal neural patterns associated with harmful knowledge, such as biological weapon development, the team can effectively switch off these functions. This technique provides a potential mechanism for developers to mitigate dual-use risks while keeping models helpful for legitimate tasks, marking a shift toward more precise safety interventions in AI development.