About Me

I started in business, not AI. My bachelor's was in Digital Business Management, and I worked in SAP infrastructure at ALDI, compliance at KPMG, and security automation at mindsquare. In 2022 I started a data science master's at Bielefeld University, which included an Erasmus year at Université Gustave Eiffel in France. As large language models became more capable, I shifted my focus toward AI safety after GPT-4 was released. During my studies I co-founded a SME (2021-2025) with two friends where we provided IT-Consulting services to small and medium-sized enterprises.

I recently completed my master's thesis work at the ELLIS Institute Tübingen under the supervision of Maksym Andriushchenko, focusing on agentic AI behaviour and alignment. That work led to the Instrumental Choices paper on measuring whether terminal-based LLM agents choose instrumental shortcuts, accepted at NeurIPS 2026. I also presented it to the red-teaming group at the UK AISI. I co-authored a study on gender stereotypes in TikTok content presented at the 2024 GOR conference. I'm doing a research internship at ExpSec through October 2026 and continuing to work on empirical AI safety. My current focus is agentic behaviour: how LLM agents act when objectives, constraints, and incentives come into tension, and how to evaluate those behaviours with clean baselines, clear metrics, and reproducible setups. In the near future I will probably focus on multi-agent systems and the challenges of coordinating them and keeping them aligned.

After the Collusion Wiki report came out, I spent some of my free time looking for more traces of AI agents on the web. I found activity on several more sites, including a teacher's chemistry wiki, as well as two compromised Hugging Face accounts and what appeared to be a training environment uploaded to RubyGems. These findings were added to the wiki's updates.

Some of this work was covered by The Washington Post, Reuters, DER SPIEGEL and Axios.