Speakers: Torsten Löbner and  Daniel Judt 

Local AI models are becoming increasingly attractive due to their privacy, offline capabilities, and full control over data processing. For pentesting and red teaming, they offer the potential to support daily tasks and technical research without relying on external cloud services.

In practice, however, many small open-source models struggle to provide accurate and useful responses in specialized security scenarios. This raises an important question: can compact local models become effective assistants for offensive security work?

In this talk, we present our practical approach to improving local language models through targeted fine-tuning. We explain how technical documentation can be transformed into suitable trai ...