Abstract or Summary:
Theory on passwords has lagged behind practice, where large providers use back-end smarts to survive with imperfect technology. Simplistic models of user and attacker behaviors have led the research community to emphasize the wrong threats. Authentication is a classification problem amenable to machine learning, with many signals in addition to the password available to largeWeb services. Passwords will continue as a useful signal for the foreseeable future, where the goal is not impregnable security but reducing harm at acceptable cost.