We are increasingly using voice models in our everyday lives - whether in chat functions for support or in automated call systems (ACD). These models are intended to fulfill a specific purpose and in some cases also access sensitive data, such as histories, behavior, processes and personal data.

The use of prompt hacking (so-called prompt injection) represents an increasing challenge in dealing with these modern language models (LLMs). Attackers can deliberately or unconsciously influence the behavior or responses of the model through targeted manipulation of text or voice input.

This and other attack possibilities as well as the potential of prompt injection will be presented in detail in the lecture.