AI regulation, agent security and innovation beyond frontier labs

a16z describes a discussion whose panelists argue that earlier technologies, from computer viruses to aviation, developed safety standards after people learned how they failed. The panelists suggest AI agents may require rethinking permissions, authentication and the security stack because they operate at enormous scale, don't get tired and probe systems in ways employees cannot. They also discuss why AI innovation may increasingly happen outside frontier labs, in software built around models.