Deepfake audio and video is already impressively accurate. Just needs to be optimized to render in realtime. Once an attacker can train an LLM to act like a target and respond to conversation in realtime, it's game over.