The company has established controls that flag when a model attempts to break out of a sandbox or successfully accesses the live internet, cordoned off its highest-risk test environments, and proposed a set of safety standards for its external testing partners, such as giving AI agents explicit instructions like “you should not access the internet.”
