A researcher-developed framework could enable attackers to conduct real-time conversations using simulated audio to compromise organizations and extract sensitive information.